The open-weights index · July 2026
Open models.
Top 33 open-source AI models
Ranked by significance| # | Model | Company | Country | Category | Size | License | Released | Why it matters |
|---|---|---|---|---|---|---|---|---|
| 01 | DeepSeek V4 Pro |  DeepSeek | China | LLM | 1.6T MoE (49B active) | MIT | 2026-04 | Current DeepSeek flagship: MIT-licensed 1M-context MoE competitive with closed frontier models. |
| 02 | DeepSeek R1 (R1-0528) | DeepSeek | China | Reasoning | 671B MoE (37B active) | MIT | 2025-01 | The open reasoning model that matched o1 and reset expectations for open-weight AI. |
| 03 | gpt-oss-120b |  OpenAI | USA | Reasoning | 117B MoE (5.1B active) | Apache 2.0 | 2025-08 | OpenAI's first open-weight LLM since GPT-2; near o4-mini reasoning on one 80GB GPU (MXFP4). |
| 04 | Qwen3.5-397B-A17B |  Alibaba | China | LLM | 397B MoE (17B active) | Apache 2.0 | 2026-02 | Alibaba's 2026 open flagship for agentic work; 201-language support, large throughput gains. |
| 05 | Kimi K2.6 |  Moonshot AI | China | LLM | 1T MoE (32B active) | Modified MIT | 2026-04 | Latest downloadable Kimi flagship; agentic model tying frontier closed models on coding. |
| 06 | Llama 4 Maverick |  Meta | USA | Multimodal | 400B MoE (17B active) | Llama 4 Community License | 2025-04 | Meta's open-weight multimodal flagship; the Llama line's final open release as Meta went proprietary. |
| 07 | GLM-5.2 |  Zhipu AI (Z.ai) | China | Reasoning | 744B MoE (~40B active) | MIT | 2026-06 | MIT-licensed 1M-context MoE trained on Huawei Ascend; rivals closed models on coding and agents. |
| 08 | Gemma 4 31B |  Google | USA | Multimodal | 31B dense (also E2B/E4B/26B MoE) | Apache 2.0 | 2026-04 | #3 open model on Arena text leaderboard; Gemma's first Apache 2.0 release, image/video input. |
| 09 | Mistral Large 3 |  Mistral AI | France | Multimodal | 675B MoE (41B active) | Apache 2.0 | 2025-12 | Mistral's flagship open MoE with image input; strongest permissively licensed European model. |
| 10 | Inkling |  Thinking Machines Lab | USA | Multimodal | 975B MoE (41B active) | Apache 2.0 | 2026-07 | Murati lab's first open model: 1M-context multimodal MoE with controllable thinking effort. |
| 11 | Nemotron 3 Ultra |  NVIDIA | USA | Reasoning | 550B MoE (55B active) | NVIDIA Open Model License | 2026-06 | Top US open-weights reasoner; ships with training data, RL environments, and recipes. |
| 12 | Kimi K2 Thinking | Moonshot AI | China | Reasoning | 1T MoE (32B active) | Modified MIT | 2025-11 | Open thinking model with native INT4 and 256K context; strongest open reasoner of late 2025. |
| 13 | Hunyuan Hy3 |  Tencent | China | LLM | 295B MoE (21B active) | Apache 2.0 | 2026-07 | Agentic-coding MoE with 256K context; Apache 2.0 with earlier regional restrictions removed. |
| 14 | MiniMax-M3 |  MiniMax | China | Multimodal | 428B MoE (~23B active) | MiniMax Community License | 2026-06 | Native multimodal 1M-context sparse-attention model; license limits large-scale commercial use. |
| 15 | Intern-S1-Pro |  Shanghai AI Lab | China | Multimodal | 1T MoE (22B active) | Apache 2.0 | 2026-02 | Largest open scientific multimodal model (512 experts); free commercial use. |
| 16 | Olmo 3.1 32B Think |  Allen Institute for AI | USA | Reasoning | 32B dense | Apache 2.0 | 2025-12 | Fully open weights, data, code, and checkpoints - the most transparent release anywhere. |
| 17 | ERNIE 4.5 (424B-A47B) |  Baidu | China | Multimodal | 424B MoE (47B active) | Apache 2.0 | 2025-06 | 10-variant open family up to 424B multimodal MoE; Baidu's flagship open release. |
| 18 | Grok 2.5 |  xAI | USA | LLM | ~270B MoE (reported) | Grok 2 Community License | 2025-08 | xAI's latest downloadable weights; the promised Grok 3 open release had not shipped by mid-2026. |
| 19 | Granite 4.1 |  IBM | USA | LLM | 3B-30B (hybrid Mamba-Transformer) | Apache 2.0 | 2026-04 | Enterprise hybrid models with 512K context; ISO 42001-certified and cryptographically signed. |
| 20 | Phi-4-reasoning-vision-15B |  Microsoft | USA | Reasoning | 15B dense | MIT | 2026-03 | Compact vision-language reasoner that decides when to think before answering. |
| 21 | Falcon H1R 7B |  Technology Innovation Institute | UAE | Reasoning | 7B (hybrid Transformer-Mamba) | Falcon TII License | 2026-01 | Best-in-class 7B reasoner per TII, beating much larger open models like Qwen3 32B. |
| 22 | Step-3.7-Flash |  StepFun | China | Multimodal | 198B MoE (~11B active) | Apache 2.0 | 2026-05 | High-efficiency sparse VLM built for real-world agents; GGUF builds available. |
| 23 | Seed-OSS-36B-Instruct |  ByteDance | China | LLM | 36B dense | Apache 2.0 | 2025-08 | Native 512K context and a user-controllable thinking budget from ByteDance Seed. |
| 24 | Qwen3-Coder-480B-A35B |  Alibaba (Qwen) | China | Coding | 480B MoE (35B active) | Apache 2.0 | 2025-07 | Flagship open agentic coding model, 256K context; Qwen3-Coder-Next 80B-A3B followed in 2026. |
| 25 | StarCoder2-15B |  BigCode (Hugging Face + ServiceNow) | International | Coding | 15B dense | BigCode OpenRAIL-M | 2024-02 | Open-science code model trained on The Stack v2 with opt-out consent; fully transparent data. |
| 26 | FLUX.2 [dev] |  Black Forest Labs | Germany | Image | 32B | FLUX.2-dev Non-Commercial License | 2025-11 | Flagship open-weight text-to-image/editing model; smaller FLUX.2 [klein] 4B is Apache 2.0. |
| 27 | ERNIE-Image | Baidu | China | Image | 8B DiT | Apache 2.0 | 2026-04 | Single-stream Diffusion Transformer; SOTA among open text-to-image, with an 8-step Turbo distill. |
| 28 | Stable Diffusion 3.5 Large |  Stability AI | UK | Image | 8.1B | Stability AI Community License | 2024-10 | The most widely forked open image family; free commercial use under $1M annual revenue. |
| 29 | Wan 2.2 (A14B) |  Alibaba | China | Video | 27B MoE (14B active) | Apache 2.0 | 2025-07 | Leading fully open video suite (T2V/I2V); later Wan versions are API-only, not open. |
| 30 | HunyuanVideo 1.5 | Tencent | China | Video | 8.3B | Tencent Hunyuan Community License | 2025-11 | Efficient successor to the 13B HunyuanVideo; top open video quality on consumer GPUs. |
| 31 | Mochi 1 |  Genmo | USA | Video | 10B | Apache 2.0 | 2024-10 | AsymmDiT text-to-video pioneer; one of the first truly permissive open video models. |
| 32 | Whisper large-v3-turbo | OpenAI | USA | Audio | 809M | MIT | 2024-10 | De facto standard open speech recognition; pruned large-v3 that runs about 8x faster. |
| 33 | Kokoro-82M v1.0 |  Hexgrad | USA | Audio | 82M | Apache 2.0 | 2025-01 | Tiny open TTS that topped the TTS Arena; 54 voices, 8 languages, runs on CPU. |
Sizes, licenses & dates verified against public releases, July 2026 · corrections: hello@opencurious.com
How to read this table.
Open-weight ≠ one license
Weights are downloadable for every row, but terms differ. Apache 2.0 and MIT allow nearly unrestricted commercial use; community licenses (Llama, Grok, MiniMax) add conditions worth reading before you ship.
How MoE sizes work
Mixture-of-experts sizes list total parameters with the active-per-token count in parentheses. A 1T MoE with 32B active runs far cheaper than a 1T dense model.
The balance of power
Chinese labs (DeepSeek, Alibaba, Moonshot, Zhipu) now ship most frontier-class open weights, while US entries range from OpenAI's gpt-oss to Thinking Machines' debut, Inkling.