LLM model comparison: releases, openness, and local memory

A model’s name tells you surprisingly little about what you can run. Llama 4 Scout activates 17 billion parameters per token but stores roughly 109 billion. GPT-6 Astra is available through hosted products, but its weights and parameter count are not public. Both facts matter more for local deployment than a leaderboard position.

This catalog compares 379 releases and size variants, from early generative transformers through releases checked on September 24, 2026. Each row links to the creator’s model card, repository, paper, or release documentation. Hugging Face links point to the publishing organization’s repository, rather than a community repackaging.

Coverage and reading conventions

The table covers major general-purpose, reasoning, multimodal, and coding families. It includes individual released sizes for the main downloadable families and significant historical releases. Base and instruction versions that share a size and release are usually represented by one row. Distinct reasoning and coding releases receive separate rows when they change the comparison.

It is a dated catalog of major releases, not a claim to enumerate every model ever uploaded. Community fine-tunes, merges, duplicate quantizations, embedding-only models, image generators, speech-only models, and every API snapshot are outside its scope. Research and restricted-access releases are labeled. Inclusion of an older model does not mean its hosted endpoint is still available.

Dates describe the named release, with preview, paper, API, and weight-release dates distinguished where they differ. A month-only date deliberately carries less precision than a day-level date. Repository creation, paper publication, first preview, and general availability can fall on different dates.

B means billion parameters. Counts are approximate published model sizes, not an exact count of every tensor in a download. Where documented, a second number gives the parameters active per token in a mixture-of-experts model. Trillion-scale counts remain in billions for comparison: 1,000B is 1T. Undisclosed counts stay undisclosed rather than repeating estimates from rumors.

The table is grouped by creator, then release date. Search this page for a family such as Kimi, Sol, Llama, or Qwen3.8. On narrow screens, scroll the table horizontally to reach the description column.

Private, open weights, and open source

Private / proprietary means the full model weights are not publicly available for local deployment. A public chatbot or API does not make its underlying model open source. “Private” here describes access to the model, not a promise about how a provider handles your prompts.

Open weights means you can obtain the trained parameters under the listed terms. Apache 2.0 and MIT are permissive open-source licenses, but a permissively licensed weight file does not by itself establish that the whole AI system meets an open-source definition. The table names the license rather than collapsing everything downloadable into a single “open source” label.

Open source is used for the Olmo releases that publish the training pipeline and data along with the model. The Open Source AI Definition considers the code and data information needed to modify the system, as well as its parameters. Other rows are not an assessment of compliance with that definition.

Custom licenses matter. Llama community terms, Gemma terms, modified MIT licenses, research licenses, and non-commercial licenses have different conditions. A license on inference code can also differ from the license on the weights. The linked model’s license file is the source for those conditions. A family name does not establish the terms for every release.

Hugging Face is both a model-hosting platform and a contributor to model research. It created the SmolLM family and participated in collaborations such as BigScience and BigCode. Hosting Qwen or Llama does not make Hugging Face the creator of those models.

How the local memory estimates work

The memory column reads 4-bit weight floor / planning budget, in decimal GB. For example, 4 / 7-10 means an idealized 4 GB of packed weights and an illustrative 7-10 GB model-process budget. These are calculated estimates, not measured hardware requirements or guarantees that a particular quantization is available.

For a model with P billion stored parameters:

Weight precision	Idealized weight storage in GB
32-bit	4 × P
16-bit, FP16 or BF16	2 × P
8-bit	P
4-bit	0.5 × P

The table’s working budget is ceil(0.6 × P + 2) to ceil(0.75 × P + 4) GB. This is a uniform planning heuristic for one inference stream, a short text context of roughly 2K-4K tokens, mostly 4-bit resident weights, and an efficient supported runtime. It allows more than packed-weight size for quantization metadata, higher-precision tensors, cache, and working buffers. It is not a vendor benchmark. Some models need more than this range even under those assumptions.

Quantization usually reduces storage by representing weights with fewer bits. It does not necessarily quantize every tensor. Supported hardware, kernels, and model architectures differ by engine. Hugging Face documents these distinctions in its quantization guide.

Context memory is separate from weights. A transformer often stores a key/value cache for the tokens it has processed. A longer prompt, more generated tokens, or more simultaneous requests can substantially increase memory. Sliding-window attention, latent attention, state-space layers, cache quantization, and offloading change the calculation. The cache documentation explains why parameter count alone cannot predict it.

For ordinary full-attention transformer layers, a useful cache estimate in bytes is 2 × layers × tokens × KV heads × head dimension × bytes per element × batch size. The factor of two accounts for keys and values. This formula does not directly describe every hybrid or latent-attention model in the table.

GPU VRAM and system RAM are different pools. A model kept entirely on a GPU needs its resident weights and working allocations in VRAM. CPU inference uses system RAM. Unified-memory machines share a pool with the operating system and other applications. CPU/GPU offloading can distribute allocations, but simply adding the two capacities does not prove a configuration will work or run quickly. Reserve additional host memory for the operating system, loading, and other software.

MoE reduces active computation, not automatically resident weight storage. The default estimates include all experts. Expert offloading can lower GPU residency, at the cost of memory transfers and a different system-RAM requirement. A 1T model with 32B active parameters is not a 32B download.

Multimodal encoders, per-layer embedding tables, speculative decoding models, and conditional-memory modules can add storage beyond a headline count. Explicit + entries identify known exclusions or architecture-specific conditions. The same caution applies to rounded size labels elsewhere. Long-context or multimodal operation needs a model-specific measurement. Training and fine-tuning memory are outside these inference estimates.

For scale, a nominal 8B model gives 4 / 7-10 GB, a 70B model gives 35 / 44-57 GB, and a 1T model gives 500 / 602-754 GB. These examples explain the arithmetic, not shopping recommendations. Before choosing hardware, check the exact checkpoint size, supported quantization, engine, intended context length, and measured peak allocation.

Large model comparison table

Every model name is a link. Local GB uses the weight-floor / planning-budget convention above. For proprietary models, local execution is unavailable regardless of how much memory a machine has. “Not established” means the public evidence checked here does not support a local deployment claim.

Model and source	Creator	Release date	Parameters, total / active	Access and license	Local GB, Q4 floor / budget	Description
Yi 34B	01.AI	2023-11	34B	Open weights: Apache 2.0	17 / 23-30	Early bilingual English/Chinese dense model.
Yi 6B	01.AI	2023-11	6B	Open weights: Apache 2.0	3 / 6-9	Early bilingual English/Chinese dense model.
Yi-1.5 34B	01.AI	2024-05	34B	Open weights: Apache 2.0	17 / 23-30	Updated Yi model with expanded coding and instruction training.
Yi-1.5 6B	01.AI	2024-05	6B	Open weights: Apache 2.0	3 / 6-9	Updated Yi model with expanded coding and instruction training.
Yi-1.5 9B	01.AI	2024-05	9B	Open weights: Apache 2.0	4.5 / 8-11	Updated Yi model with expanded coding and instruction training.
OLMo 7B	Ai2	2024-02	7B	Open source: Apache 2.0	3.5 / 7-10	Open research family publishing weights, training data, code, and intermediate artifacts.
OLMo 2 13B	Ai2	2024-11	13B	Open source: Apache 2.0	6.5 / 10-14	Open research family publishing weights, training data, code, and intermediate artifacts.
OLMo 2 7B	Ai2	2024-11	7B	Open source: Apache 2.0	3.5 / 7-10	Open research family publishing weights, training data, code, and intermediate artifacts.
OLMo 2 32B	Ai2	2025-03	32B	Open source: Apache 2.0	16 / 22-28	Open research family publishing weights, training data, code, and intermediate artifacts.
Olmo 3 7B	Ai2	2025-11-20	7B	Open source: Apache 2.0	3.5 / 7-10	Open research family publishing weights, training data, code, and intermediate artifacts.
Olmo 3 Think 32B	Ai2	2025-11-20	32B	Open source: Apache 2.0	16 / 22-28	Open research family publishing weights, training data, code, and intermediate artifacts.
Olmo 3.1 Think 32B	Ai2	2025-12	32B	Open source: Apache 2.0	16 / 22-28	Open research family publishing weights, training data, code, and intermediate artifacts.
Jamba v0.1	AI21 Labs	2024-03	52B / 12B active	Open weights: Apache 2.0	26 / 34-43	Hybrid Mamba/Transformer MoE for long-context text.
Jamba 1.5 Large	AI21 Labs	2024-08-22	398B / 94B active	Open weights: Jamba license	199 / 241-303	Hybrid Mamba/Transformer MoE for long-context text.
Jamba 1.5 Mini	AI21 Labs	2024-08-22	52B / 12B active	Open weights: Jamba license	26 / 34-43	Hybrid Mamba/Transformer MoE for long-context text.
Qwen 7B	Alibaba	2023-08	7B	Open weights: Tongyi Qianwen	3.5 / 7-10	Original bilingual Qwen family, with sizes released in stages.
Qwen 14B	Alibaba	2023-09	14B	Open weights: Tongyi Qianwen	7 / 11-15	Original bilingual Qwen family, with sizes released in stages.
Qwen 1.8B	Alibaba	2023-11	1.8B	Open weights: Tongyi Qianwen	0.9 / 4-6	Original bilingual Qwen family, with sizes released in stages.
Qwen 72B	Alibaba	2023-11	72B	Open weights: Tongyi Qianwen	36 / 46-58	Original bilingual Qwen family, with sizes released in stages.
Qwen1.5 0.5B	Alibaba	2024-02	0.5B	Open weights: Tongyi Qianwen research	0.25 / 3-5	Expanded dense Qwen family, with sizes released in stages.
Qwen1.5 1.8B	Alibaba	2024-02	1.8B	Open weights: Tongyi Qianwen research	0.9 / 4-6	Expanded dense Qwen family, with sizes released in stages.
Qwen1.5 14B	Alibaba	2024-02	14B	Open weights: Tongyi Qianwen	7 / 11-15	Expanded dense Qwen family, with sizes released in stages.
Qwen1.5 4B	Alibaba	2024-02	4B	Open weights: Tongyi Qianwen research	2 / 5-7	Expanded dense Qwen family, with sizes released in stages.
Qwen1.5 72B	Alibaba	2024-02	72B	Open weights: Tongyi Qianwen	36 / 46-58	Expanded dense Qwen family, with sizes released in stages.
Qwen1.5 7B	Alibaba	2024-02	7B	Open weights: Tongyi Qianwen	3.5 / 7-10	Expanded dense Qwen family, with sizes released in stages.
Qwen1.5 110B	Alibaba	2024-04	110B	Open weights: Tongyi Qianwen	55 / 68-87	Expanded dense Qwen family, with sizes released in stages.
Qwen1.5 32B	Alibaba	2024-04	32B	Open weights: Tongyi Qianwen research	16 / 22-28	Expanded dense Qwen family, with sizes released in stages.
Qwen2 0.5B	Alibaba	2024-06-07	0.5B	Open weights: Apache 2.0	0.25 / 3-5	Second-generation multilingual Qwen dense model.
Qwen2 1.5B	Alibaba	2024-06-07	1.5B	Open weights: Apache 2.0	0.75 / 3-6	Second-generation multilingual Qwen dense model.
Qwen2 57B-A14B	Alibaba	2024-06-07	57B / 14B active	Open weights: Apache 2.0	28.5 / 37-47	Sparse MoE member of the Qwen2 family.
Qwen2 72B	Alibaba	2024-06-07	72B	Open weights: Qwen license	36 / 46-58	Largest dense Qwen2 model.
Qwen2 7B	Alibaba	2024-06-07	7B	Open weights: Apache 2.0	3.5 / 7-10	Second-generation multilingual Qwen dense model.
Qwen2.5-Coder 1.5B	Alibaba	2024-09	1.5B	Open weights: Apache 2.0	0.75 / 3-6	Code-specialized Qwen2.5 family, released in stages.
Qwen2.5-Coder 7B	Alibaba	2024-09	7B	Open weights: Apache 2.0	3.5 / 7-10	Code-specialized Qwen2.5 family, released in stages.
Qwen2.5 0.5B	Alibaba	2024-09-19	0.5B	Open weights: Apache 2.0	0.25 / 3-5	Qwen2.5 dense model for multilingual text, coding, and structured outputs.
Qwen2.5 1.5B	Alibaba	2024-09-19	1.5B	Open weights: Apache 2.0	0.75 / 3-6	Qwen2.5 dense model for multilingual text, coding, and structured outputs.
Qwen2.5 14B	Alibaba	2024-09-19	14B	Open weights: Apache 2.0	7 / 11-15	Qwen2.5 dense model for multilingual text, coding, and structured outputs.
Qwen2.5 32B	Alibaba	2024-09-19	32B	Open weights: Apache 2.0	16 / 22-28	Qwen2.5 dense model for multilingual text, coding, and structured outputs.
Qwen2.5 3B	Alibaba	2024-09-19	3B	Open weights: Qwen research	1.5 / 4-7	Qwen2.5 dense model for multilingual text, coding, and structured outputs.
Qwen2.5 72B	Alibaba	2024-09-19	72B	Open weights: Qwen license	36 / 46-58	Qwen2.5 dense model for multilingual text, coding, and structured outputs.
Qwen2.5 7B	Alibaba	2024-09-19	7B	Open weights: Apache 2.0	3.5 / 7-10	Qwen2.5 dense model for multilingual text, coding, and structured outputs.
Qwen2.5-Coder 0.5B	Alibaba	2024-11	0.5B	Open weights: Apache 2.0	0.25 / 3-5	Code-specialized Qwen2.5 family, released in stages.
Qwen2.5-Coder 14B	Alibaba	2024-11	14B	Open weights: Apache 2.0	7 / 11-15	Code-specialized Qwen2.5 family, released in stages.
Qwen2.5-Coder 32B	Alibaba	2024-11	32B	Open weights: Apache 2.0	16 / 22-28	Code-specialized Qwen2.5 family, released in stages.
Qwen2.5-Coder 3B	Alibaba	2024-11	3B	Open weights: Qwen research	1.5 / 4-7	Code-specialized Qwen2.5 family, released in stages.
Qwen2.5-VL 3B	Alibaba	2025-01	3B	Open weights: Apache 2.0	1.5 / 4-7	Visual understanding model for images, documents, and video.
Qwen2.5-VL 72B	Alibaba	2025-01	72B	Open weights: Qwen license	36 / 46-58	Visual understanding model for images, documents, and video.
Qwen2.5-VL 7B	Alibaba	2025-01	7B	Open weights: Apache 2.0	3.5 / 7-10	Visual understanding model for images, documents, and video.
Qwen2.5-VL 32B	Alibaba	2025-03	32B	Open weights: Apache 2.0	16 / 22-28	Visual understanding model for images, documents, and video.
Qwen3 0.6B	Alibaba	2025-04-29	0.6B	Open weights: Apache 2.0	0.3 / 3-5	Dense Qwen3 model supporting thinking and non-thinking modes.
Qwen3 1.7B	Alibaba	2025-04-29	1.7B	Open weights: Apache 2.0	0.85 / 4-6	Dense Qwen3 model supporting thinking and non-thinking modes.
Qwen3 14B	Alibaba	2025-04-29	14B	Open weights: Apache 2.0	7 / 11-15	Dense Qwen3 model supporting thinking and non-thinking modes.
Qwen3 235B-A22B	Alibaba	2025-04-29	235B / 22B active	Open weights: Apache 2.0	117.5 / 143-181	Sparse Qwen3 MoE supporting thinking and non-thinking modes.
Qwen3 30B-A3B	Alibaba	2025-04-29	30B / 3B active	Open weights: Apache 2.0	15 / 20-27	Sparse Qwen3 MoE supporting thinking and non-thinking modes.
Qwen3 32B	Alibaba	2025-04-29	32B	Open weights: Apache 2.0	16 / 22-28	Dense Qwen3 model supporting thinking and non-thinking modes.
Qwen3 4B	Alibaba	2025-04-29	4B	Open weights: Apache 2.0	2 / 5-7	Dense Qwen3 model supporting thinking and non-thinking modes.
Qwen3 8B	Alibaba	2025-04-29	8B	Open weights: Apache 2.0	4 / 7-10	Dense Qwen3 model supporting thinking and non-thinking modes.
Qwen3-Coder 30B-A3B	Alibaba	2025-07	30B / 3B active	Open weights: Apache 2.0	15 / 20-27	Smaller coding MoE for more accessible local deployment.
Qwen3-Coder 480B-A35B	Alibaba	2025-07	480B / 35B active	Open weights: Apache 2.0	240 / 290-364	Large coding MoE aimed at software agents.
Qwen3-VL 235B-A22B	Alibaba	2025-09	235B / 22B active	Open weights: Apache 2.0	117.5 / 143-181	MoE Qwen3 vision-language model.
Qwen3-Next 80B-A3B	Alibaba	2025-09-11	80B / 3B active	Open weights: Apache 2.0	40 / 50-64	Hybrid-attention, highly sparse MoE.
Qwen3-VL 2B	Alibaba	2025-10	2B	Open weights: Apache 2.0	1 / 4-6	Dense Qwen3 vision-language model, including document and video understanding.
Qwen3-VL 30B-A3B	Alibaba	2025-10	30B / 3B active	Open weights: Apache 2.0	15 / 20-27	MoE Qwen3 vision-language model.
Qwen3-VL 32B	Alibaba	2025-10	32B	Open weights: Apache 2.0	16 / 22-28	Dense Qwen3 vision-language model, including document and video understanding.
Qwen3-VL 4B	Alibaba	2025-10	4B	Open weights: Apache 2.0	2 / 5-7	Dense Qwen3 vision-language model, including document and video understanding.
Qwen3-VL 8B	Alibaba	2025-10	8B	Open weights: Apache 2.0	4 / 7-10	Dense Qwen3 vision-language model, including document and video understanding.
Qwen3-Coder-Next	Alibaba	2026-02	80B / 3B active	Open weights: Apache 2.0	40 / 50-64	Coding model based on the Qwen3-Next architecture.
Qwen3.5-397B-A17B	Alibaba	2026-02-16	397B / 17B active	Open weights: Apache 2.0	198.5 / 241-302	Unified vision-language Qwen3.5 model with hybrid attention.
Qwen3.5-122B-A10B	Alibaba	2026-02-24	122B / 10B active	Open weights: Apache 2.0	61 / 76-96	Unified vision-language Qwen3.5 model with hybrid attention.
Qwen3.5-27B	Alibaba	2026-02-24	27B	Open weights: Apache 2.0	13.5 / 19-25	Unified vision-language Qwen3.5 model with hybrid attention.
Qwen3.5-35B-A3B	Alibaba	2026-02-24	35B / 3B active	Open weights: Apache 2.0	17.5 / 23-31	Unified vision-language Qwen3.5 model with hybrid attention.
Qwen3.5-0.8B	Alibaba	2026-03-02	0.8B	Open weights: Apache 2.0	0.4 / 3-5	Unified vision-language Qwen3.5 model with hybrid attention.
Qwen3.5-2B	Alibaba	2026-03-02	2B	Open weights: Apache 2.0	1 / 4-6	Unified vision-language Qwen3.5 model with hybrid attention.
Qwen3.5-4B	Alibaba	2026-03-02	4B	Open weights: Apache 2.0	2 / 5-7	Unified vision-language Qwen3.5 model with hybrid attention.
Qwen3.5-9B	Alibaba	2026-03-02	9B	Open weights: Apache 2.0	4.5 / 8-11	Unified vision-language Qwen3.5 model with hybrid attention.
Qwen3.6-35B-A3B	Alibaba	2026-04-16	35B / 3B active	Open weights: Apache 2.0	17.5 / 23-31	Updated Qwen model emphasizing coding and multimodal reasoning.
Qwen3.6-27B	Alibaba	2026-04-22	27B	Open weights: Apache 2.0	13.5 / 19-25	Updated Qwen model emphasizing coding and multimodal reasoning.
Qwen3.8 2.4T-A95B	Alibaba	2026-08-12	2400B / 95B active	Open weights: Qwen3.8-Max license	1200 / 1442-1804	Downloadable Qwen-Max-class MoE.
Qwen3.8 27B	Alibaba	2026-08-14	27B	Open weights: Apache 2.0	13.5 / 19-25	Dense multimodal Qwen3.8 model for smaller deployments.
Nova Lite	Amazon	2024-12-03	Undisclosed	Private / proprietary	Not available locally	Low-cost multimodal hosted model.
Nova Micro	Amazon	2024-12-03	Undisclosed	Private / proprietary	Not available locally	Text-only lightweight hosted model.
Nova Pro	Amazon	2024-12-03	Undisclosed	Private / proprietary	Not available locally	Larger multimodal hosted model.
Nova Premier	Amazon	2025-04	Undisclosed	Private / proprietary	Not available locally	Largest first-generation Nova model.
Nova 2 Lite	Amazon	2025-12-02	Undisclosed	Private / proprietary	Not available locally	Reasoning model for efficient agent and multimodal tasks.
Claude 1	Anthropic	2023-03-14	Undisclosed	Private / proprietary	Not available locally	Original Claude assistant model.
Claude 2	Anthropic	2023-07-11	Undisclosed	Private / proprietary	Not available locally	Long-context successor to Claude 1.
Claude Instant 1.2	Anthropic	2023-08-09	Undisclosed	Private / proprietary	Not available locally	Early lightweight Claude model.
Claude 2.1	Anthropic	2023-11-21	Undisclosed	Private / proprietary	Not available locally	Expanded context and early tool-use support.
Claude 3 Opus	Anthropic	2024-03-04	Undisclosed	Private / proprietary	Not available locally	Largest original Claude 3 tier, with vision.
Claude 3 Sonnet	Anthropic	2024-03-04	Undisclosed	Private / proprietary	Not available locally	Middle Claude 3 tier, with vision.
Claude 3 Haiku	Anthropic	2024-03-13	Undisclosed	Private / proprietary	Not available locally	Smallest Claude 3 tier for fast responses.
Claude 3.5 Sonnet	Anthropic	2024-06-20	Undisclosed	Private / proprietary	Not available locally	Improved coding and visual understanding.
Claude 3.5 Sonnet v2	Anthropic	2024-10-22	Undisclosed	Private / proprietary	Not available locally	Updated Sonnet accompanying the computer-use beta.
Claude 3.5 Haiku	Anthropic	2024-11	Undisclosed	Private / proprietary	Not available locally	Smaller Claude 3.5 model announced in October.
Claude 3.7 Sonnet	Anthropic	2025-02-24	Undisclosed	Private / proprietary	Not available locally	Introduced optional extended thinking.
Claude Opus 4	Anthropic	2025-05-22	Undisclosed	Private / proprietary	Not available locally	Claude 4 flagship for sustained coding and reasoning.
Claude Sonnet 4	Anthropic	2025-05-22	Undisclosed	Private / proprietary	Not available locally	Balanced Claude 4 model for everyday work and coding.
Claude Opus 4.1	Anthropic	2025-08-05	Undisclosed	Private / proprietary	Not available locally	Opus update focused on agentic tasks and coding.
Claude Sonnet 4.5	Anthropic	2025-09-29	Undisclosed	Private / proprietary	Not available locally	Sonnet update focused on coding agents.
Claude Haiku 4.5	Anthropic	2025-10-15	Undisclosed	Private / proprietary	Not available locally	Small, fast model in the Claude 4.5 generation.
Claude Opus 4.5	Anthropic	2025-11-24	Undisclosed	Private / proprietary	Not available locally	Opus update for coding, agents, and computer use.
Claude Opus 4.6	Anthropic	2026-02-05	Undisclosed	Private / proprietary	Not available locally	Opus reasoning and long-context update.
Claude Sonnet 4.6	Anthropic	2026-02-17	Undisclosed	Private / proprietary	Not available locally	Sonnet update for coding and professional work.
Claude Mythos Preview	Anthropic	2026-04 (restricted)	Undisclosed	Private / proprietary	Not available locally	Restricted research preview for selected defensive security partners.
Claude Opus 4.7	Anthropic	2026-04-16	Undisclosed	Private / proprietary	Not available locally	Improved software engineering, verification, and image understanding.
Claude Opus 4.8	Anthropic	2026-05-28	Undisclosed	Private / proprietary	Not available locally	Updated Opus with more efficient tool use and collaboration.
Claude Fable 5	Anthropic	2026-06-09	Undisclosed	Private / proprietary	Not available locally	General-use Claude 5 model. Access was suspended and restored during rollout.
Claude Mythos 5	Anthropic	2026-06-09 (restricted)	Undisclosed	Private / proprietary	Not available locally	Restricted-access Claude 5 model with advanced cyber capabilities.
Claude Sonnet 5	Anthropic	2026-06-30	Undisclosed	Private / proprietary	Not available locally	Sonnet model emphasizing planning and autonomous tool use.
Claude Opus 5	Anthropic	2026-07-24	Undisclosed	Private / proprietary	Not available locally	Claude 5 model aimed at everyday coding and professional work.
ERNIE 4.5 0.3B	Baidu	2025-06-30 (weights)	0.3B	Open weights: Apache 2.0	0.15 / 3-5	Baidu downloadable model. VL variants accept visual inputs.
ERNIE 4.5 21B-A3B	Baidu	2025-06-30 (weights)	21B / 3B active	Open weights: Apache 2.0	10.5 / 15-20	Baidu downloadable model. VL variants accept visual inputs.
ERNIE 4.5 300B-A47B	Baidu	2025-06-30 (weights)	300B / 47B active	Open weights: Apache 2.0	150 / 182-229	Baidu downloadable model. VL variants accept visual inputs.
ERNIE 4.5 VL-28B-A3B	Baidu	2025-06-30 (weights)	28B / 3B active	Open weights: Apache 2.0	14 / 19-25	Baidu downloadable model. VL variants accept visual inputs.
ERNIE 4.5 VL-424B-A47B	Baidu	2025-06-30 (weights)	424B / 47B active	Open weights: Apache 2.0	212 / 257-322	Baidu downloadable model. VL variants accept visual inputs.
StarCoder	BigCode (Hugging Face and partners)	2023-05	15.5B	Open weights: BigCode OpenRAIL-M	7.75 / 12-16	Code-completion model trained by BigCode.
StarCoder2 15B	BigCode (Hugging Face and partners)	2024-02	15B	Open weights: BigCode OpenRAIL-M	7.5 / 11-16	Code model trained on The Stack v2.
StarCoder2 3B	BigCode (Hugging Face and partners)	2024-02	3B	Open weights: BigCode OpenRAIL-M	1.5 / 4-7	Smaller StarCoder2 code-completion model.
StarCoder2 7B	BigCode (Hugging Face and partners)	2024-02	7B	Open weights: BigCode OpenRAIL-M	3.5 / 7-10	Smaller StarCoder2 code-completion model.
BLOOM	BigScience	2022-07-12	176B	Open weights: BigScience RAIL	88 / 108-136	Multilingual research model from the BigScience collaboration.
Command R	Cohere / Cohere For AI	2024-03	35B	Open weights: CC-BY-NC 4.0	17.5 / 23-31	Enterprise model emphasizing retrieval, grounded answers, and tool use.
Command R+	Cohere / Cohere For AI	2024-04	104B	Open weights: CC-BY-NC 4.0	52 / 65-82	Enterprise model emphasizing retrieval, grounded answers, and tool use.
Aya Expanse 32B	Cohere / Cohere For AI	2024-10	32B	Open weights: CC-BY-NC 4.0	16 / 22-28	Multilingual instruction model.
Aya Expanse 8B	Cohere / Cohere For AI	2024-10	8B	Open weights: CC-BY-NC 4.0	4 / 7-10	Multilingual instruction model.
Command A	Cohere / Cohere For AI	2025-03-13	111B	Open weights: CC-BY-NC 4.0	55.5 / 69-88	Enterprise model emphasizing retrieval, grounded answers, and tool use.
DBRX	Databricks	2024-03-27	132B / 36B active	Open weights: Databricks Open Model	66 / 82-103	Fine-grained MoE for general text and coding.
DeepSeek Coder 1.3B	DeepSeek	2023-11	1.3B	Open weights: DeepSeek model license	0.65 / 3-5	Early code model trained for completion and repository context.
DeepSeek Coder 33B	DeepSeek	2023-11	33B	Open weights: DeepSeek model license	16.5 / 22-29	Early code model trained for completion and repository context.
DeepSeek Coder 6.7B	DeepSeek	2023-11	6.7B	Open weights: DeepSeek model license	3.35 / 7-10	Early code model trained for completion and repository context.
DeepSeek LLM 67B	DeepSeek	2023-11	67B	Open weights: DeepSeek model license	33.5 / 43-55	Early general-purpose dense DeepSeek model.
DeepSeek LLM 7B	DeepSeek	2023-11	7B	Open weights: DeepSeek model license	3.5 / 7-10	Early general-purpose dense DeepSeek model.
DeepSeek-V2	DeepSeek	2024-05	236B / 21B active	Open weights: DeepSeek model license	118 / 144-181	MoE model introducing Multi-head Latent Attention.
DeepSeek-V2-Lite	DeepSeek	2024-05	16B / 2.4B active	Open weights: DeepSeek model license	8 / 12-16	Small V2 MoE for experimentation and local inference.
DeepSeek-Coder-V2	DeepSeek	2024-06	236B / 21B active	Open weights: DeepSeek model license	118 / 144-181	Code-specialized V2 with expanded language coverage.
DeepSeek-Coder-V2-Lite	DeepSeek	2024-06	16B / 2.4B active	Open weights: DeepSeek model license	8 / 12-16	Smaller code-specialized V2 MoE.
DeepSeek-V2.5	DeepSeek	2024-09	236B / 21B active	Open weights: DeepSeek model license	118 / 144-181	Combined general chat and coding capabilities.
DeepSeek-V3	DeepSeek	2024-12-26	671B / 37B active	Open weights: DeepSeek model license	335.5 / 405-508	Large MoE with FP8 training and multi-token prediction.
DeepSeek-R1	DeepSeek	2025-01-20	671B / 37B active	Open weights: MIT	335.5 / 405-508	Reasoning model trained with reinforcement learning and cold-start data.
DeepSeek-R1-Distill-Llama-70B	DeepSeek	2025-01-20	70B	Open weights: MIT + base terms	35 / 44-57	Smaller dense student trained on R1 outputs. Original Qwen or Llama terms also matter.
DeepSeek-R1-Distill-Llama-8B	DeepSeek	2025-01-20	8B	Open weights: MIT + base terms	4 / 7-10	Smaller dense student trained on R1 outputs. Original Qwen or Llama terms also matter.
DeepSeek-R1-Distill-Qwen-1.5B	DeepSeek	2025-01-20	1.5B	Open weights: MIT + base terms	0.75 / 3-6	Smaller dense student trained on R1 outputs. Original Qwen or Llama terms also matter.
DeepSeek-R1-Distill-Qwen-14B	DeepSeek	2025-01-20	14B	Open weights: MIT + base terms	7 / 11-15	Smaller dense student trained on R1 outputs. Original Qwen or Llama terms also matter.
DeepSeek-R1-Distill-Qwen-32B	DeepSeek	2025-01-20	32B	Open weights: MIT + base terms	16 / 22-28	Smaller dense student trained on R1 outputs. Original Qwen or Llama terms also matter.
DeepSeek-R1-Distill-Qwen-7B	DeepSeek	2025-01-20	7B	Open weights: MIT + base terms	3.5 / 7-10	Smaller dense student trained on R1 outputs. Original Qwen or Llama terms also matter.
DeepSeek-R1-Zero	DeepSeek	2025-01-20	671B / 37B active	Open weights: MIT	335.5 / 405-508	Research reasoning model trained with reinforcement learning without an initial supervised stage.
DeepSeek-V3-0324	DeepSeek	2025-03-24	671B / 37B active	Open weights: MIT	335.5 / 405-508	V3 update with improved reasoning and instruction following.
DeepSeek-R1-0528	DeepSeek	2025-05-28	671B / 37B active	Open weights: MIT	335.5 / 405-508	Updated R1 reasoning checkpoint.
DeepSeek-V3.1	DeepSeek	2025-08	671B / 37B active	Open weights: MIT	335.5 / 405-508	Hybrid thinking and non-thinking successor to V3.
DeepSeek-V3.2	DeepSeek	2025-12	671B / 37B active	Open weights: MIT	335.5 / 405-508	Sparse-attention model combining reasoning and tool use.
DeepSeek-V4 Flash	DeepSeek	2026-04-24 (preview)	284B / 13B active	Open weights: MIT	142 / 173-217	Smaller V4 MoE supporting a million-token context.
DeepSeek-V4 Pro	DeepSeek	2026-04-24 (preview)	1600B / 49B active	Open weights: MIT	800 / 962-1204	Largest V4 model. General availability followed in August.
DeepSeek-V4.1 Flash	DeepSeek	2026-09-10	748B / 8 prefill / 16 decodeB active	Open weights: MIT	374+ / 451-565+ (conditional)	552B backbone plus 196B Engram memory. Estimate covers both, excludes uncounted auxiliary modules.
GPT-Neo 1.3B	EleutherAI	2021-03	1.3B	Open weights: MIT	0.65 / 3-5	Early community-trained autoregressive transformer.
GPT-Neo 2.7B	EleutherAI	2021-03	2.7B	Open weights: MIT	1.35 / 4-7	Early community-trained autoregressive transformer.
GPT-J	EleutherAI	2021-06	6B	Open weights: Apache 2.0	3 / 6-9	Early openly released GPT-style dense text model.
GPT-NeoX	EleutherAI	2022-02	20B	Open weights: Apache 2.0	10 / 14-19	Large open research model trained with the GPT-NeoX framework.
Pythia 12B	EleutherAI	2023-02	12B	Open weights: Apache 2.0	6 / 10-13	Largest Pythia research model, with intermediate training checkpoints.
T5 11B	Google	2019-10 (paper)	11B	Open weights: Apache 2.0	5.5 / 9-13	Encoder-decoder research model casting NLP tasks as text-to-text generation.
LaMDA	Google	2022-01 (paper)	137B	Private / proprietary	Not available locally	Dialogue-focused research model. Largest published configuration.
PaLM	Google	2022-04 (paper)	540B	Private / proprietary	Not available locally	Large dense model trained with the Pathways system. Largest published configuration.
FLAN-T5 XXL	Google	2022-10	11B	Open weights: Apache 2.0	5.5 / 9-13	Instruction-tuned T5 model. Encoder-decoder caching differs from decoder-only models.
PaLM 2	Google	2023-05	Undisclosed	Private / proprietary	Not available locally	Multilingual successor to PaLM used in early Bard deployments.
Gemini 1.0 Pro	Google DeepMind	2023-12-06	Undisclosed	Private / proprietary	Not available locally	First Gemini Pro release for general multimodal tasks.
Gemini 1.0 Ultra	Google DeepMind	2024-02	Undisclosed	Private / proprietary	Not available locally	Largest first-generation Gemini tier.
Gemini 1.5 Pro	Google DeepMind	2024-02-15 (preview)	Undisclosed	Private / proprietary	Not available locally	Long-context MoE generation.
Gemma 2B	Google DeepMind	2024-02-21	2B	Open weights: Gemma terms	1 / 4-6	First downloadable Gemma text models.
Gemma 7B	Google DeepMind	2024-02-21	7B	Open weights: Gemma terms	3.5 / 7-10	First downloadable Gemma text models.
Gemini 1.5 Flash	Google DeepMind	2024-05 (preview)	Undisclosed	Private / proprietary	Not available locally	Faster, lower-cost sibling to 1.5 Pro.
Gemma 2 27B	Google DeepMind	2024-06-27	27B	Open weights: Gemma terms	13.5 / 19-25	Second-generation Gemma dense text model.
Gemma 2 9B	Google DeepMind	2024-06-27	9B	Open weights: Gemma terms	4.5 / 8-11	Second-generation Gemma dense text model.
Gemma 2 2B	Google DeepMind	2024-07-31	2B	Open weights: Gemma terms	1 / 4-6	Second-generation Gemma dense text model.
Gemini 2.0 Flash	Google DeepMind	2024-12 (experimental)	Undisclosed	Private / proprietary	Not available locally	Second-generation Flash model with expanded tool and multimodal capabilities.
Gemini 2.0 Flash-Lite	Google DeepMind	2025-02 (preview)	Undisclosed	Private / proprietary	Not available locally	Lower-cost Gemini 2.0 tier.
Gemma 3 12B	Google DeepMind	2025-03-12	12B	Open weights: Gemma terms	6 / 10-13	Gemma 3 multimodal model with image inputs.
Gemma 3 1B	Google DeepMind	2025-03-12	1B	Open weights: Gemma terms	0.5 / 3-5	Gemma 3 text model.
Gemma 3 27B	Google DeepMind	2025-03-12	27B	Open weights: Gemma terms	13.5 / 19-25	Gemma 3 multimodal model with image inputs.
Gemma 3 4B	Google DeepMind	2025-03-12	4B	Open weights: Gemma terms	2 / 5-7	Gemma 3 multimodal model with image inputs.
Gemini 2.5 Pro	Google DeepMind	2025-03-25 (experimental)	Undisclosed	Private / proprietary	Not available locally	Thinking model for reasoning and coding.
Gemini 2.5 Flash	Google DeepMind	2025-04 (preview)	Undisclosed	Private / proprietary	Not available locally	Reasoning model balancing latency and capability.
Gemini 2.5 Flash-Lite	Google DeepMind	2025-06 (preview)	Undisclosed	Private / proprietary	Not available locally	Smallest hosted Gemini 2.5 tier.
Gemma 3n E2B	Google DeepMind	2025-06-26 (full release)	5B	Open weights: Gemma terms	2.5 / 5-8	Multimodal edge model. Effective size differs from total storage. Selective loading changes RAM use.
Gemma 3n E4B	Google DeepMind	2025-06-26 (full release)	8B	Open weights: Gemma terms	4 / 7-10	Multimodal edge model. Effective size differs from total storage. Selective loading changes RAM use.
Gemma 3 270M	Google DeepMind	2025-08	0.27B	Open weights: Gemma terms	0.135 / 3-5	Tiny text model for specialized on-device tasks.
Gemini 3 Pro	Google DeepMind	2025-11 (preview)	Undisclosed	Private / proprietary	Not available locally	Gemini 3 flagship reasoning and multimodal model.
Gemini 3 Flash	Google DeepMind	2025-12	Undisclosed	Private / proprietary	Not available locally	Faster model in the Gemini 3 generation.
Gemini 3.1 Pro	Google DeepMind	2026-02-19 (preview)	Undisclosed	Private / proprietary	Not available locally	Updated Pro model for difficult reasoning tasks.
Gemini 3.1 Flash-Lite	Google DeepMind	2026-03	Undisclosed	Private / proprietary	Not available locally	Low-latency Gemini 3.1 tier.
Gemma 4 26B-A4B	Google DeepMind	2026-03-31	25.2B / 3.8B active	Open weights: Apache 2.0	12.6+ / 18-24+ (encoder)	Sparse multimodal Gemma 4. Add approximately 0.55B vision parameters.
Gemma 4 31B	Google DeepMind	2026-03-31	30.7B	Open weights: Apache 2.0	15.35+ / 21-28+ (encoder)	Dense multimodal Gemma 4. Add approximately 0.55B vision parameters.
Gemma 4 E2B	Google DeepMind	2026-03-31	5.1B	Open weights: Apache 2.0	2.55+ / 6-8+ (encoders)	Effective-size label excludes large per-layer embeddings. Count includes embeddings, add encoder memory.
Gemma 4 E4B	Google DeepMind	2026-03-31	8B	Open weights: Apache 2.0	4+ / 7-10+ (encoders)	Effective-size label excludes large per-layer embeddings. Count includes embeddings, add encoder memory.
Gemini 3.5 Flash	Google DeepMind	2026-05-19	Undisclosed	Private / proprietary	Not available locally	Flash update for reasoning and agent workflows.
Gemma 4 12B Unified	Google DeepMind	2026-06-03	11.95B	Open weights: Apache 2.0	5.975 / 10-13	Encoder-free model accepting text, image, and audio inputs.
Gemini 3.5 Flash-Lite	Google DeepMind	2026-07-21	Undisclosed	Private / proprietary	Not available locally	Efficient model for high-volume multimodal work.
Gemini 3.6 Flash	Google DeepMind	2026-07-21	Undisclosed	Private / proprietary	Not available locally	Successor to 3.5 Flash.
Gemini 3.7 Flash	Google DeepMind	2026-08-13	Undisclosed	Private / proprietary	Not available locally	Intermediate Flash update for coding and reasoning.
Gemini 3.8 Flash	Google DeepMind	2026-09-02	Undisclosed	Private / proprietary	Not available locally	Flash update focused on agentic reasoning and coding.
SmolLM2 0.135B	Hugging Face	2024-11	0.135B	Open weights: Apache 2.0	0.0675 / 3-5	Small text models with published training resources.
SmolLM2 0.36B	Hugging Face	2024-11	0.36B	Open weights: Apache 2.0	0.18 / 3-5	Small text models with published training resources.
SmolLM2 1.7B	Hugging Face	2024-11	1.7B	Open weights: Apache 2.0	0.85 / 4-6	Small text models with published training resources.
SmolLM3 3B	Hugging Face	2025-07-08	3B	Open weights: Apache 2.0	1.5 / 4-7	Hugging Face text model with reasoning, long context, and multilingual support.
Granite 3.0 2B	IBM	2024-10	2B	Open weights: Apache 2.0	1 / 4-6	Enterprise text model with instruction following and tool support.
Granite 3.0 8B	IBM	2024-10	8B	Open weights: Apache 2.0	4 / 7-10	Enterprise text model with instruction following and tool support.
Granite 4.0 H Small	IBM	2025-10-02	32B / 9B active	Open weights: Apache 2.0	16 / 22-28	Hybrid Mamba/Transformer MoE for enterprise tasks.
LongCat-Flash-Chat	Meituan	2025-09-01	560B / 18.6-31.3B active	Open weights: MIT	280 / 338-424	Dynamic MoE varying active compute by token.
LongCat-Flash-Thinking-2601	Meituan	2026-01	560B / 18.6-31.3B active	Open weights: MIT	280 / 338-424	Reasoning-tuned LongCat for tool-using tasks.
OPT 175B	Meta	2022-05	175B	Restricted research weights: OPT	87.5 / 107-136 (if access granted)	Meta research model released with training logs and access restrictions.
LLaMA 1 13B	Meta	2023-02	13B	Open weights: research-only	6.5 / 10-14	Original LLaMA release with restricted research access.
LLaMA 1 33B	Meta	2023-02	33B	Open weights: research-only	16.5 / 22-29	Original LLaMA release with restricted research access.
LLaMA 1 65B	Meta	2023-02	65B	Open weights: research-only	32.5 / 41-53	Original LLaMA release with restricted research access.
LLaMA 1 7B	Meta	2023-02	7B	Open weights: research-only	3.5 / 7-10	Original LLaMA release with restricted research access.
Llama 2 13B	Meta	2023-07-18	13B	Open weights: Llama community	6.5 / 10-14	Second-generation text model. Chat variants were released alongside base models.
Llama 2 70B	Meta	2023-07-18	70B	Open weights: Llama community	35 / 44-57	Second-generation text model. Chat variants were released alongside base models.
Llama 2 7B	Meta	2023-07-18	7B	Open weights: Llama community	3.5 / 7-10	Second-generation text model. Chat variants were released alongside base models.
Code Llama 13B	Meta	2023-08	13B	Open weights: Llama community	6.5 / 10-14	Code-focused Llama 2 derivative, with separate Python and instruction variants.
Code Llama 34B	Meta	2023-08	34B	Open weights: Llama community	17 / 23-30	Code-focused Llama 2 derivative, with separate Python and instruction variants.
Code Llama 7B	Meta	2023-08	7B	Open weights: Llama community	3.5 / 7-10	Code-focused Llama 2 derivative, with separate Python and instruction variants.
Code Llama 70B	Meta	2024-01	70B	Open weights: Llama community	35 / 44-57	Larger Code Llama release for code generation.
Llama 3 70B	Meta	2024-04-18	70B	Open weights: Llama community	35 / 44-57	Third-generation text model with a larger vocabulary and training corpus.
Llama 3 8B	Meta	2024-04-18	8B	Open weights: Llama community	4 / 7-10	Third-generation text model with a larger vocabulary and training corpus.
Llama 3.1 405B	Meta	2024-07-23	405B	Open weights: Llama community	202.5 / 245-308	Llama generation supporting longer contexts and multilingual use.
Llama 3.1 70B	Meta	2024-07-23	70B	Open weights: Llama community	35 / 44-57	Llama generation supporting longer contexts and multilingual use.
Llama 3.1 8B	Meta	2024-07-23	8B	Open weights: Llama community	4 / 7-10	Llama generation supporting longer contexts and multilingual use.
Llama 3.2 1B	Meta	2024-09-25	1B	Open weights: Llama community	0.5 / 3-5	Compact text model intended for local and edge use.
Llama 3.2 3B	Meta	2024-09-25	3B	Open weights: Llama community	1.5 / 4-7	Compact text model intended for local and edge use.
Llama 3.2 Vision 11B	Meta	2024-09-25	11B	Open weights: Llama community	5.5 / 9-13	Vision-capable Llama combining image understanding with text generation.
Llama 3.2 Vision 90B	Meta	2024-09-25	90B	Open weights: Llama community	45 / 56-72	Vision-capable Llama combining image understanding with text generation.
Llama 3.3 70B	Meta	2024-12	70B	Open weights: Llama community	35 / 44-57	Instruction-tuned 70B refresh emphasizing multilingual quality.
Llama 4 Maverick	Meta	2025-04-05	400B / 17B active	Open weights: Llama 4 community	200 / 242-304	Native multimodal MoE. Memory follows the total count, not its 17B active count.
Llama 4 Scout	Meta	2025-04-05	109B / 17B active	Open weights: Llama 4 community	54.5 / 68-86	Native multimodal MoE. Memory follows the total count, not its 17B active count.
Muse Spark	Meta Superintelligence Labs	2026-04-08	Undisclosed	Private / proprietary	Not available locally	Hosted multimodal reasoning model in the Muse family.
Muse Spark 1.1	Meta Superintelligence Labs	2026-07-09	Undisclosed	Private / proprietary	Not available locally	Hosted multimodal reasoning model in the Muse family.
Muse Glimmer 30B	Meta Superintelligence Labs	2026-08-10	30B	Open weights: Apache 2.0	15 / 20-27	Multimodal model distilled for local agents. Optional speculative decoder needs additional memory.
Muse Spark 1.3	Meta Superintelligence Labs	2026-09-02	Undisclosed	Private / proprietary	Not available locally	Hosted multimodal reasoning model in the Muse family.
Phi-1	Microsoft	2023-06	1.3B	Open weights: MIT	0.65 / 3-5	Small code model trained with curated and synthetic data.
Phi-1.5	Microsoft	2023-09	1.3B	Open weights: MIT	0.65 / 3-5	Small text model exploring data quality over scale.
Phi-2	Microsoft	2023-12	2.7B	Open weights: MIT	1.35 / 4-7	Compact model trained on curated educational-style data.
Phi-3 mini	Microsoft	2024-04	3.8B	Open weights: MIT	1.9 / 5-7	Small instruction model for constrained deployments.
Phi-3 medium	Microsoft	2024-05	14B	Open weights: MIT	7 / 11-15	Largest original dense Phi-3 model.
Phi-3 small	Microsoft	2024-05	7B	Open weights: MIT	3.5 / 7-10	Middle Phi-3 dense model.
Phi-3.5 MoE	Microsoft	2024-08	41.9B / 6.6B active	Open weights: MIT	20.95 / 28-36	Sparse Phi model with 6.6B active parameters.
Phi-3.5 mini	Microsoft	2024-08	3.8B	Open weights: MIT	1.9 / 5-7	Updated compact Phi instruction model.
Phi-4	Microsoft	2024-12-12	14B	Open weights: MIT	7 / 11-15	Small dense model emphasizing reasoning and synthetic training data.
Phi-4 mini	Microsoft	2025-02	3.8B	Open weights: MIT	1.9 / 5-7	Smaller Phi-4 instruction model.
Phi-4 multimodal	Microsoft	2025-02	5.6B	Open weights: MIT	2.8 / 6-9	Compact model accepting text, image, and audio.
Phi-4 reasoning	Microsoft	2025-04	14B	Open weights: MIT	7 / 11-15	Reasoning-tuned Phi-4 model.
MiniMax-Text-01	MiniMax	2025-01	456B / 45.9B active	Open weights: MiniMax model license	228 / 276-346	Hybrid linear-attention MoE for long contexts.
MiniMax-M1	MiniMax	2025-06	456B / 45.9B active	Open weights: Apache 2.0	228 / 276-346	Long-context reasoning model based on MiniMax-Text-01.
MiniMax-M2	MiniMax	2025-10	230B / 10B active	Open weights: modified MIT	115 / 140-177	Sparse model focused on coding and agents.
MiniMax-M2.1	MiniMax	2025-12	230B / 10B active	Open weights: modified MIT	115 / 140-177	M2 update targeting multilingual software development.
MiniMax-M2.5	MiniMax	2026-02	230B / 10B active	Open weights: modified MIT	115 / 140-177	M2 generation optimized for coding and office tasks.
MiniMax-M2.7	MiniMax	2026-03 (API), 2026-04 (weights)	230B / 10B active	Open weights: non-commercial	115 / 140-177	M2 update for software engineering. Commercial use requires separate authorization.
MiniMax-M3	MiniMax	2026-06-01	428B / 23B active	Open weights: MiniMax community	214 / 259-325	Native multimodal model using MiniMax Sparse Attention.
Mistral 7B	Mistral AI	2023-09-27	7.3B	Open weights: Apache 2.0	3.65 / 7-10	Compact dense model using grouped-query and sliding-window attention.
Mixtral 8x7B	Mistral AI	2023-12	46.7B / 12.9B active	Open weights: Apache 2.0	23.35 / 31-40	Sparse MoE. Eight experts do not mean eight complete 7B models.
Mistral Large (original)	Mistral AI	2024-02	Undisclosed	Private / proprietary	Not available locally	Original hosted flagship. Do not reuse Large 2 parameter counts.
Mixtral 8x22B	Mistral AI	2024-04	141B / 39B active	Open weights: Apache 2.0	70.5 / 87-110	Larger sparse Mixtral model.
Codestral 22B	Mistral AI	2024-05	22B	Open weights: Mistral non-production	11 / 16-21	Code model with fill-in-the-middle support and restricted production use.
Mistral NeMo	Mistral AI	2024-07	12B	Open weights: Apache 2.0	6 / 10-13	Dense model co-developed with NVIDIA.
Mistral Large 2	Mistral AI	2024-07-24	123B	Open weights: Mistral research	61.5 / 76-97	Large dense model. Commercial self-hosting uses separate terms.
Pixtral 12B	Mistral AI	2024-09	12B	Open weights: Apache 2.0	6 / 10-13	Multimodal model combining a language backbone with a vision encoder.
Pixtral Large	Mistral AI	2024-11	124B	Open weights: Mistral research	62 / 77-97	Large multimodal model with a 123B language backbone.
Mistral Small 3	Mistral AI	2025-01-30	24B	Open weights: Apache 2.0	12 / 17-22	Dense instruction model designed for efficient deployment.
Mistral Small 3.1	Mistral AI	2025-03	24B	Open weights: Apache 2.0	12 / 17-22	Small model adding vision and a longer context.
Devstral Small	Mistral AI	2025-05	24B	Open weights: Apache 2.0	12 / 17-22	Software engineering model co-developed with All Hands AI.
Mistral Medium 3	Mistral AI	2025-05	Undisclosed	Private / proprietary	Not available locally	Hosted model for professional tasks.
Magistral Small	Mistral AI	2025-06	24B	Open weights: Apache 2.0	12 / 17-22	Downloadable reasoning model.
Devstral 2	Mistral AI	2025-12	123B	Open weights: modified MIT	61.5 / 76-97	Large software engineering model.
Devstral Small 2	Mistral AI	2025-12	24B	Open weights: Apache 2.0	12 / 17-22	Smaller software engineering model with vision.
Ministral 3 14B	Mistral AI	2025-12-02	14B	Open weights: Apache 2.0	7 / 11-15	Small dense multimodal member of the Mistral 3 family.
Ministral 3 3B	Mistral AI	2025-12-02	3B	Open weights: Apache 2.0	1.5 / 4-7	Small dense multimodal member of the Mistral 3 family.
Ministral 3 8B	Mistral AI	2025-12-02	8B	Open weights: Apache 2.0	4 / 7-10	Small dense multimodal member of the Mistral 3 family.
Mistral Large 3	Mistral AI	2025-12-02	675B / 41B active	Open weights: Apache 2.0	337.5 / 407-511	Large multimodal MoE, replacing the dense Large 2 architecture.
Mistral Small 4	Mistral AI	2026-03-16	119B / 6B active	Open weights: Apache 2.0	59.5 / 74-94	MoE unifying reasoning, coding, and multimodal inputs.
Mistral Medium 3.5	Mistral AI	2026-04-29	128B	Open weights: modified MIT	64 / 79-100	Dense multimodal model with adjustable reasoning. License has large-company exceptions.
Kimi K1.5	Moonshot AI	2025-01-20 (report)	Undisclosed	Private / proprietary	Not available locally	Multimodal reasoning research model. Full weights were not released.
Kimi-VL	Moonshot AI	2025-04-10	16B / 2.8B active	Open weights: MIT	8 / 12-16	Small multimodal MoE with vision understanding.
Kimi-Dev 72B	Moonshot AI	2025-06-17	72B	Open weights: MIT	36 / 46-58	Software engineering model trained for repository-level issue resolution.
Kimi K2	Moonshot AI	2025-07-11	1000B / 32B active	Open weights: modified MIT	500 / 602-754	Trillion-parameter text MoE emphasizing coding and tool use.
Kimi K2 0905	Moonshot AI	2025-09-05	1000B / 32B active	Open weights: modified MIT	500 / 602-754	K2 instruction update with longer context.
Kimi K2 Thinking	Moonshot AI	2025-11-06	1000B / 32B active	Open weights: modified MIT	500 / 602-754	Reasoning variant for extended tool-using tasks.
Kimi K2.5	Moonshot AI	2026-01-27	1000B / 32B active	Open weights: modified MIT	500 / 602-754	Native multimodal K2 generation for visual coding and agent work.
Kimi K2.6	Moonshot AI	2026-04-20	1000B / 32B active	Open weights: modified MIT	500 / 602-754	K2 update for longer coding tasks and visual understanding.
Kimi K2.7 Code	Moonshot AI	2026-06-12	1000B / 32B active	Open weights: modified MIT	500 / 602-754	Coding-focused K2 release with vision and extended agent training.
Kimi K3	Moonshot AI	2026-07-16 (API), 07-27 (weights)	2800B / 104B active	Open weights: Kimi K3 license	1400 / 1682-2104	Large multimodal MoE with Kimi Delta Attention and Attention Residuals.
Nemotron 3 Nano	NVIDIA	2025-12	30B / 3B active	Open weights: NVIDIA Nemotron license	15 / 20-27	Hybrid Mamba, attention, and MoE model for reasoning and agents.
Nemotron 3 Super	NVIDIA	2026-03-11	120B / 12B active	Open weights: NVIDIA Nemotron license	60 / 74-94	Hybrid Mamba, attention, and MoE model for reasoning and agents.
Nemotron 3.5 Lightning	NVIDIA	2026-08-11	30B / 3B active	Open weights: OpenMDW 1.1	15 / 20-27	Hybrid Mamba, attention, and MoE model for reasoning and agents.
GPT (original)	OpenAI	2018-06	0.117B	Open weights: MIT	0.0585 / 3-5	Early generative pretraining model, followed by task-specific fine-tuning.
GPT-2 0.124B	OpenAI	2019-02	0.124B	Open weights: MIT	0.062 / 3-5	Staged release of a general text-completion model. Largest weights released in November.
GPT-2 0.355B	OpenAI	2019-05	0.355B	Open weights: MIT	0.1775 / 3-5	Staged release of a general text-completion model. Largest weights released in November.
GPT-2 0.774B	OpenAI	2019-08	0.774B	Open weights: MIT	0.387 / 3-5	Staged release of a general text-completion model. Largest weights released in November.
GPT-2 1.5B	OpenAI	2019-11	1.5B	Open weights: MIT	0.75 / 3-6	Staged release of a general text-completion model. Largest weights released in November.
GPT-3	OpenAI	2020-05 (paper)	175B	Private / proprietary	Not available locally	Few-shot prompting milestone. Count refers to the largest model in the paper.
GPT-3.5 Turbo	OpenAI	2023-03-01	Undisclosed	Private / proprietary	Not available locally	Chat-oriented API model from the early ChatGPT era.
GPT-4	OpenAI	2023-03-14	Undisclosed	Private / proprietary	Not available locally	Stronger instruction following and reasoning than the GPT-3.5 generation.
GPT-4 Turbo	OpenAI	2023-11-06 (preview)	Undisclosed	Private / proprietary	Not available locally	Longer-context GPT-4 variant. Vision GA followed in April 2024.
GPT-4o	OpenAI	2024-05-13	Undisclosed	Private / proprietary	Not available locally	Omni generation combining text and image understanding with broader multimodal capabilities.
GPT-4o mini	OpenAI	2024-07-18	Undisclosed	Private / proprietary	Not available locally	Smaller hosted model for inexpensive text and vision tasks.
o1-mini	OpenAI	2024-09-12	Undisclosed	Private / proprietary	Not available locally	Smaller reasoning model focused on math and coding.
o1-preview	OpenAI	2024-09-12	Undisclosed	Private / proprietary	Not available locally	Early reasoning model that spends additional inference time solving problems.
o1	OpenAI	2024-12	Undisclosed	Private / proprietary	Not available locally	Production successor to o1-preview.
o3-mini	OpenAI	2025-01-31	Undisclosed	Private / proprietary	Not available locally	Compact reasoning model with configurable reasoning effort.
GPT-4.5	OpenAI	2025-02-27 (preview)	Undisclosed	Private / proprietary	Not available locally	Research preview focused on conversational knowledge and writing.
GPT-4.1	OpenAI	2025-04-14	Undisclosed	Private / proprietary	Not available locally	General model emphasizing coding, instruction following, and long contexts.
GPT-4.1 mini	OpenAI	2025-04-14	Undisclosed	Private / proprietary	Not available locally	Smaller GPT-4.1 tier for lower-cost applications.
GPT-4.1 nano	OpenAI	2025-04-14	Undisclosed	Private / proprietary	Not available locally	Smallest GPT-4.1 tier for lightweight tasks.
o3	OpenAI	2025-04-16	Undisclosed	Private / proprietary	Not available locally	Reasoning model supporting tools and visual problem solving.
o4-mini	OpenAI	2025-04-16	Undisclosed	Private / proprietary	Not available locally	Efficient reasoning model for math, coding, and visual tasks.
o3-pro	OpenAI	2025-06-10	Undisclosed	Private / proprietary	Not available locally	Higher-compute o3 offering for difficult problems.
gpt-oss-120b	OpenAI	2025-08-05	117B / 5.1B active	Open weights: Apache 2.0	58.5 / 73-92	Downloadable reasoning MoE with native MXFP4 expert weights.
gpt-oss-20b	OpenAI	2025-08-05	21B / 3.6B active	Open weights: Apache 2.0	10.5 / 15-20	Downloadable reasoning MoE with native MXFP4 expert weights.
GPT-5	OpenAI	2025-08-07	Undisclosed	Private / proprietary	Not available locally	Reasoning generation for coding and general tool use.
GPT-5 mini	OpenAI	2025-08-07	Undisclosed	Private / proprietary	Not available locally	Smaller GPT-5 model for clearly specified tasks.
GPT-5 nano	OpenAI	2025-08-07	Undisclosed	Private / proprietary	Not available locally	Lowest-cost original GPT-5 tier for classification and summarization.
GPT-5-Codex	OpenAI	2025-09	Undisclosed	Private / proprietary	Not available locally	GPT-5 variant trained for software engineering agents.
GPT-5 Pro	OpenAI	2025-10-06 (API)	Undisclosed	Private / proprietary	Not available locally	Higher-compute GPT-5 offering.
GPT-5.1	OpenAI	2025-11	Undisclosed	Private / proprietary	Not available locally	Updated general reasoning model with adaptive task effort.
GPT-5.1-Codex	OpenAI	2025-11	Undisclosed	Private / proprietary	Not available locally	Coding-focused GPT-5.1 variant.
GPT-5.1-Codex Max	OpenAI	2025-11	Undisclosed	Private / proprietary	Not available locally	Coding variant designed for longer software engineering work.
GPT-5.1-Codex mini	OpenAI	2025-11	Undisclosed	Private / proprietary	Not available locally	Smaller coding-focused GPT-5.1 variant.
GPT-5.2	OpenAI	2025-12	Undisclosed	Private / proprietary	Not available locally	Successor targeting professional knowledge work and coding.
GPT-5.2 Pro	OpenAI	2025-12	Undisclosed	Private / proprietary	Not available locally	Higher-compute GPT-5.2 offering.
GPT-5.2-Codex	OpenAI	2025-12	Undisclosed	Private / proprietary	Not available locally	GPT-5.2 variant for agentic software engineering.
GPT-5.3-Codex	OpenAI	2026-02	Undisclosed	Private / proprietary	Not available locally	Coding and computer-use model for extended agent work.
GPT-5.3 Instant	OpenAI	2026-03	Undisclosed	Private / proprietary	Not available locally	Conversational model exposed through a chat-oriented alias.
GPT-5.4	OpenAI	2026-03-05	Undisclosed	Private / proprietary	Not available locally	General professional model combining reasoning, coding, and computer use.
GPT-5.4 Pro	OpenAI	2026-03-05	Undisclosed	Private / proprietary	Not available locally	Higher-compute GPT-5.4 offering.
GPT-5.4 mini	OpenAI	2026-03-17	Undisclosed	Private / proprietary	Not available locally	Smaller GPT-5.4 tier for higher-volume work.
GPT-5.4 nano	OpenAI	2026-03-17	Undisclosed	Private / proprietary	Not available locally	Lightweight GPT-5.4 tier for simple tasks.
GPT-5.5	OpenAI	2026-04-24 (API)	Undisclosed	Private / proprietary	Not available locally	Professional reasoning model with text and image inputs.
GPT-5.5 Pro	OpenAI	2026-04-24 (API)	Undisclosed	Private / proprietary	Not available locally	Higher-compute GPT-5.5 offering.
GPT-5.6 Luna	OpenAI	2026-07-09 (GA)	Undisclosed	Private / proprietary	Not available locally	Smallest GPT-5.6 tier for fast, high-volume work.
GPT-5.6 Sol	OpenAI	2026-07-09 (GA)	Undisclosed	Private / proprietary	Not available locally	Flagship GPT-5.6 tier for complex professional tasks.
GPT-5.6 Terra	OpenAI	2026-07-09 (GA)	Undisclosed	Private / proprietary	Not available locally	Middle GPT-5.6 tier balancing capability and cost.
GPT-6 Astra	OpenAI	2026-09-03	Undisclosed	Private / proprietary	Not available locally	Higher-capability GPT-6 tier for complex reasoning, research, coding, and computer use.
GPT-6 Luna	OpenAI	2026-09-22	Undisclosed	Private / proprietary	Not available locally	Lower-cost GPT-6 reasoning model accepting text and image inputs.
GPT-6 Sol	OpenAI	2026-09-22	Undisclosed	Private / proprietary	Not available locally	GPT-6 reasoning model accepting text and image inputs.
InternLM2 20B	Shanghai AI Laboratory	2024-01	20B	Open weights: InternLM license	10 / 14-19	Bilingual long-context model.
InternLM2 7B	Shanghai AI Laboratory	2024-01	7B	Open weights: InternLM license	3.5 / 7-10	Bilingual long-context model.
Step 3.5 Flash	StepFun	2026-02	196.81B / 11B active	Open weights: Apache 2.0	98.405 / 121-152	Sparse MoE emphasizing reasoning, coding, and agent tasks.
Hunyuan-Large	Tencent	2024-11	389B / 52B active	Open weights: Tencent Hunyuan	194.5 / 236-296	Large bilingual MoE.
Hunyuan-A13B	Tencent	2025-06-27 (weights)	80B / 13B active	Open weights: Tencent Hunyuan	40 / 50-64	Smaller hybrid-thinking MoE.
Falcon 40B	TII	2023-05	40B	Open weights: Apache 2.0	20 / 26-34	Larger first-generation Falcon text model.
Falcon 7B	TII	2023-05	7B	Open weights: Apache 2.0	3.5 / 7-10	TII text model trained on RefinedWeb.
Falcon 180B	TII	2023-09-06	180B	Open weights: Falcon 180B TII license	90 / 110-139	Large dense Falcon with custom use and hosting terms.
Falcon 3 10B	TII	2024-12	10B	Open weights: Falcon-LLM license	5 / 8-12	Compact Falcon generation for efficient text inference.
Falcon 3 1B	TII	2024-12	1B	Open weights: Falcon-LLM license	0.5 / 3-5	Compact Falcon generation for efficient text inference.
Falcon 3 3B	TII	2024-12	3B	Open weights: Falcon-LLM license	1.5 / 4-7	Compact Falcon generation for efficient text inference.
Falcon 3 7B	TII	2024-12	7B	Open weights: Falcon-LLM license	3.5 / 7-10	Compact Falcon generation for efficient text inference.
Falcon-H1 0.5B	TII	2025-05-21	0.5B	Open weights: Falcon-LLM license	0.25 / 3-5	Hybrid state-space and attention model.
Falcon-H1 1.5B	TII	2025-05-21	1.5B	Open weights: Falcon-LLM license	0.75 / 3-6	Hybrid state-space and attention model.
Falcon-H1 34B	TII	2025-05-21	34B	Open weights: Falcon-LLM license	17 / 23-30	Hybrid state-space and attention model.
Falcon-H1 3B	TII	2025-05-21	3B	Open weights: Falcon-LLM license	1.5 / 4-7	Hybrid state-space and attention model.
Falcon-H1 7B	TII	2025-05-21	7B	Open weights: Falcon-LLM license	3.5 / 7-10	Hybrid state-space and attention model.
Grok-1 (weights)	xAI	2024-03-17	314B	Open weights: Apache 2.0	157 / 191-240	Base MoE weights, distinct from the hosted Grok assistant.
Grok-1.5	xAI / SpaceXAI	2024-03	Undisclosed	Private / proprietary	Not available locally	Hosted successor with longer context and improved reasoning.
Grok-2	xAI / SpaceXAI	2024-08 (hosted), 2025-08 (weights)	Undisclosed	Open weights: Grok 2 community	Q4 unknown. Published recipe: 8 GPUs, each >40 GB	Grok-2 weights released after hosted launch. Publisher provides an FP8 deployment recipe.
Grok-3	xAI / SpaceXAI	2025-02	Undisclosed	Private / proprietary	Not available locally	Grok generation introducing stronger reasoning modes.
Grok-3 mini	xAI / SpaceXAI	2025-02	Undisclosed	Private / proprietary	Not available locally	Smaller hosted reasoning model.
Grok-4	xAI / SpaceXAI	2025-07-09	Undisclosed	Private / proprietary	Not available locally	Reasoning model integrated with live search and tools.
Grok-4 Fast	xAI / SpaceXAI	2025-09	Undisclosed	Private / proprietary	Not available locally	Efficient hosted model with reasoning and non-reasoning modes.
Grok-4.1	xAI / SpaceXAI	2025-11	Undisclosed	Private / proprietary	Not available locally	Grok update focused on conversation and reliability.
Grok-4.20	xAI / SpaceXAI	2026-03-10 (API)	Undisclosed	Private / proprietary	Not available locally	Hosted Grok release alongside a multi-agent offering.
Grok-4.5	xAI / SpaceXAI	2026-07-08 (API)	Undisclosed	Private / proprietary	Not available locally	Hosted model emphasizing coding, agents, and knowledge work.
Grok-4.6	xAI / SpaceXAI	2026-08-12 (API)	Undisclosed	Private / proprietary	Not available locally	Hosted coding and reasoning model with text and image inputs.
ChatGLM-6B	Z.ai / Zhipu AI	2023-03	6B	Open weights: ChatGLM model license	3 / 6-9	Early bilingual chat model.
ChatGLM2-6B	Z.ai / Zhipu AI	2023-06	6B	Open weights: ChatGLM model license	3 / 6-9	Updated bilingual model with longer context.
ChatGLM3-6B	Z.ai / Zhipu AI	2023-10	6B	Open weights: ChatGLM model license	3 / 6-9	ChatGLM generation adding tool-use capabilities.
GLM-4-9B	Z.ai / Zhipu AI	2024-06	9B	Open weights: GLM-4 license	4.5 / 8-11	Compact multilingual instruction model.
GLM-4.5	Z.ai / Zhipu AI	2025-07	355B / 32B active	Open weights: MIT	177.5 / 215-271	Large MoE for reasoning, coding, and agents.
GLM-4.5-Air	Z.ai / Zhipu AI	2025-07	106B / 12B active	Open weights: MIT	53 / 66-84	Smaller GLM-4.5 MoE.
GLM-4.6	Z.ai / Zhipu AI	2025-09	355B / 32B active	Open weights: MIT	177.5 / 215-271	GLM update for longer coding and tool-use tasks.
GLM-4.7	Z.ai / Zhipu AI	2025-12	355B / 32B active	Open weights: MIT	177.5 / 215-271	GLM update emphasizing coding and agent behavior.
GLM-5	Z.ai / Zhipu AI	2026-02	744B / 40B active	Open weights: MIT	372 / 449-562	Larger sparse-attention MoE for systems engineering.
GLM-5.1	Z.ai / Zhipu AI	2026-04	744B / 40B active	Open weights: MIT	372 / 449-562	Post-training update for extended engineering tasks.
GLM-5.2	Z.ai / Zhipu AI	2026-06	744B / 40B active	Open weights: MIT	372 / 449-562	GLM generation with a million-token context and revised attention indexing.
GLM-5.3	Z.ai / Zhipu AI	2026-08	744B / 40B active	Open weights: GLM-5.3 license	372 / 449-562	Post-training successor using the GLM-5.2 base. License differs from 5.2.
GLM-5.3-Flash	Z.ai / Zhipu AI	2026-08	320B / 18B active	Open weights: MIT	160 / 194-244	Native multimodal GLM with sparse and linear attention.

What the comparison tells you

Astra, Sol, and Luna are model tiers within OpenAI’s named generations. They are not separate labs. GPT-5.6 Sol and GPT-6 Sol are different releases, while changing the reasoning effort of one model does not establish a new parameter count. Google’s Project Astra is also a different use of the word Astra, so a bare name is not enough to identify a model.

Llama and Gemma illustrate why licenses belong on individual rows. Their generations do not all use the same terms. The same is true of GLM and MiniMax updates. The existence of older open weights does not imply that a provider’s newest hosted model is downloadable.

A local deployment decision starts with the exact checkpoint and your memory budget. A quality comparison starts with the task, evaluation method, and inference settings. More total parameters can increase storage without producing a proportional improvement on your task. More active parameters can increase computation without revealing the training data or post-training quality.

Descriptions identify what each release was designed to do. They do not rank all models on a common benchmark. For a meaningful quality comparison, hold the prompts, tool access, reasoning budget, and evaluation criteria constant. See evaluation methodology for the differences between a model score and the performance of a complete agent system.

References

The model links in the table are the primary references for individual entries. Release histories help distinguish a model launch from a later update to its card: OpenAI’s API changelog, DeepSeek’s release log, Kimi’s research archive, Meta’s Llama release table, Qwen’s release history, and Gemma’s release notes.

For deployment, the linked model cards and their license files take precedence over a family label. The Hugging Face quantization documentation and cache guide support the memory discussion. The numeric planning range is this page’s stated heuristic, not a requirement published by those sources.

Related topics

• LLM reasoning benchmarks and metrics: how model capabilities are evaluated.
• Model distillation: how smaller models learn from larger ones.
• Natural language processing: the wider field surrounding language models.
• Context window management: fitting useful information into a model’s context.