Three things to remember
- Only 23 of the 37 open-weight models are Apache-2.0 or MIT; the other 14 need clause-by-clause reading of community or research licences.
- Twelve of the top 20 are open-weight: Kimi K3 at 95.3 (#2), GLM-5.3-Flash 93.7 (#6), DeepSeek-V4.1-Flash 93.4 (#7) and Qwen3.8-VL-Flash 93.2 (#8).
- The 24 new entries cover general, document, agent, edge, embodied and generation work; generation models rank low on the composite because they are lopsided, not because they are weak.
The seventh edition is not a routine backfill. We ran the same eight-dimension scorecard over a board that grew from 35 to 56 models: 24 entries added, three removed, and downloadable-weight models up from 12 to 37 — 66% of the list. It is the first time this site has more open-weight models than closed APIs.
This essay answers three questions: who the 37 are, where they sit on the score curve, and what “open weight” actually counts as here.
The totals first
| Metric | Sixth edition | Seventh edition |
|---|---|---|
| Models | 35 | 56 |
| Score points | 280 | 448 |
| Open weights | 12 | 37 |
| Evidence A / B / C | 15 / 18 / 2 | 22 / 32 / 2 |
| Essays | 28 | 31 |
The evidence mix moved too: grade A rises from 15 to 22, grade B from 18 to 32. The new entries include models with complete weight cards, licences and evaluation scripts, and models supported mainly by technical reports and community reproduction. We did not treat “has a GitHub repository” as grade A, and we did not add points for being open.
How the 37 split by licence
“Open weights” is not “OSI open source”. The openSource field here means one thing only: the weights can be downloaded publicly. The actual conditions live in the licence field. Roughly, the 37 fall into four groups:
| Licence family | Models | Examples |
|---|---|---|
| Apache-2.0 | 18 | Qwen3.8-VL-Flash, Qwen3.8-27B, Qwen3-Omni-30B, Intern-S2, PaddleOCR-VL 1.6, Wan2.2 |
| MIT | 5 | DeepSeek-V4.1-Flash, GLM-5.3-Flash, Ling-3.0-flash-VL, DeepSeek-OCR 2 (MIT-family) |
| Community licence | 10 | Kimi K3, MiniMax M3, MiMo-V2.6-Pro, Step 5 Preview, GLM-Image, HunyuanImage 3.0 |
| Research licence | 4 | LongCat-Next, Emu3.5, BAGEL-2, Zidong Taichu 2.0 |
Only the first two groups are broadly permissive for commercial use and redistribution: 23 models. The other 14 need clause-by-clause reading. MiniCPM-V 4.5 is labelled Apache-2.0 but the vendor asks for commercial registration, so its licence cell says so.
Open weights have reached tier S
Twelve of the top 20 are open-weight models:
| Rank | Model | Composite | Licence |
|---|---|---|---|
| #2 | Kimi K3 | 95.3 | Community |
| #6 | GLM-5.3-Flash | 93.7 | MIT |
| #7 | DeepSeek-V4.1-Flash | 93.4 | MIT |
| #8 | Qwen3.8-VL-Flash | 93.2 | Apache-2.0 |
| #9 | Step 5 Preview | 93.0 | StepFun community |
| #11 | MiMo-V2.6-Pro | 91.4 | Xiaomi community |
| #12 | MiniMax M3 | 91.1 | MiniMax community |
| #13 | Qwen3.8-27B | 90.5 | Apache-2.0 |
| #16 | Step 3.7 Flash | 89.2 | StepFun community |
| #17 | Qwen3-Omni-30B | 88.9 | Apache-2.0 |
| #19 | DeepSeek-V4-Flash-Vision-Exp | 87.4 | MIT |
| #20 | dots3-note preview | 87.3 | Apache-2.0 |
The change versus the sixth edition is not “one more open model”. Open weights now occupy several top-ten slots: #2, #6, #7, #8 and #9 are all downloadable. Kimi K3 is second overall at 95.3, 0.1 behind Qwen3.8-Max; GLM-5.3-Flash is #6 under MIT. For teams that need on-prem deployment, fine-tuning or cost control, the shortlist can now be built entirely inside the top ten.
The 24 new entries
| Rank | Model | Score | Licence | Role |
|---|---|---|---|---|
| #11 | MiMo-V2.6-Pro | 91.4 | Xiaomi community | 1T-A42B omni, 1M context |
| #12 | MiniMax M3 | 91.1 | MiniMax community | 428B-A23B, 1M context, agent 93 |
| #13 | Qwen3.8-27B | 90.5 | Apache-2.0 | 27B dense, single-machine deployment |
| #16 | Step 3.7 Flash | 89.2 | StepFun community | 196B-A11B, ¥2/M output |
| #19 | DeepSeek-V4-Flash-Vision-Exp | 87.4 | MIT | 256K document-vision experiment |
| #20 | dots3-note preview | 87.3 | Apache-2.0 | Xiaohongshu notes and voice input |
| #22 | Ling-3.0-flash-VL | 85.9 | MIT | Chinese documents and finance |
| #26 | Intern-S2-Preview-397B | 84.4 | Apache-2.0 | Scientific charts and paper layouts |
| #29 | MiniCPM-V 4.7 | 81.8 | Apache-2.0 | Edge image-text assistant |
| #33 | Zidong Taichu 5.0-9B | 80.6 | Apache-2.0 | Private baseline for government and education |
| #35 | Step3-VL-10B | 80.4 | Apache-2.0 | Single-GPU starter base |
| #38 | Qwen-UI-Agent | 79.5 | Apache-2.0 | GUI agent specialist, agent 98 |
| #39 | SenseNova U1.5 Lite | 78.8 | Apache-2.0 | 8B unified understanding and generation |
| #45 | MiniCPM-o 4.5 | 76.7 | Apache-2.0 | 9B edge omni model |
| #47 | InternVL-U 4B | 75.8 | Apache-2.0 | 4B unified understand, generate, edit |
| #48 | HunyuanImage 3.0 | 74.3 | Tencent Hunyuan community | 80B image generation |
| #49 | PaddleOCR-VL 1.6 | 71.5 | Apache-2.0 | 0.9B document parsing, document 98 |
| #50 | RoboBrain 2.5 | 70.6 | Apache-2.0 | Open embodied brain |
| #51 | GLM-Image | 70.6 | Zhipu community | Image model trained on Ascend |
| #52 | SenseNova-SI-1.3-8B | 69.8 | Apache-2.0 | Spatial-intelligence small model |
| #53 | MiniMax H3 | 68.1 | MiniMax community | Omni generation |
| #54 | Kairos 3.0-4B | 66.3 | Apache-2.0 | 4B world model |
| #55 | Wan2.2 | 65.1 | Apache-2.0 | Video-generation base |
| #56 | HunyuanVideo 1.5 | 63.7 | Tencent Hunyuan community | 8.3B video generation |
A low rank is not a bad model. Wan2.2 and HunyuanVideo 1.5 score 94 and 95 on generation, but are weak on documents, agents and Chinese, so the composite drags them to the bottom. This board rewards “no hole”, not “the longest plank”. If you only need generation, read the generation score and unit cost, not the composite.
Three new signals
One, there is a new agent leader. Qwen-UI-Agent scores 98 on agents, ahead of Kimi K3 at 96, on the back of MobileWorld-Real 92.2 and AndroidDaily 97.5. Its composite is only 79.5 and it ranks #38 — the definition of a specialist: strong at phone and desktop operation, weak at general understanding.
Two, documents now split 99/98. DeepSeek-OCR 2 stays first at 99; PaddleOCR-VL 1.6 is alone in second at 98, in 0.9B parameters, and runs on CPU. The cost curve for document processing is about to be redrawn by these two.
Three, generation models sink together. Emu3.5 scores 98 on generation, HunyuanImage 3.0 and MiniMax H3 score 96, HunyuanVideo 1.5 95, Wan2.2 and GLM-Image 94 — yet they rank #34, #48, #53, #56, #55 and #51. Keeping them on one board shows the cost of being lopsided; it does not deny what they do well.
Counting rules and limits
- The 37 are “weights downloadable”, not “OSI-certified”. Of those, 23 are Apache-2.0 or MIT and 14 are community or research licences.
- Evidence grade is independent of openness. Kimi K3 is grade A; GLM-5.3-Flash is grade B. Grade B does not mean unusable, only that independent reproduction is thinner.
- The composite, weights and same-band rule are unchanged. The sample changed: 56 models, 448 score points.
- Sample-dependent derived statistics still use the sixth-edition basis. The correlation matrix, PCA, price-band counts, sycophancy and dialect experiment means need a 56-model recomputation and are queued for the next pass.
How to use the table
For on-prem, read the licence before the rank: Kimi K3 (community), GLM-5.3-Flash and DeepSeek-V4.1-Flash (MIT), and Qwen3.8-VL-Flash or Qwen3.8-27B (Apache-2.0) represent three different legal costs.
For documents, go straight to DeepSeek-OCR 2 and PaddleOCR-VL 1.6; do not let the composite decide.
For agents, first decide whether you need GUI operation or general tool calling. For GUI, look at Qwen-UI-Agent; for general work, Kimi K3, GLM-5.3-Flash and MiMo-V2.6-Pro remain the shortlist.
For a general open-weight workhorse, #2, #6, #8 and #13 are four answers at different prices and deployment shapes; #11 and #12 suit teams that need long context and omni input.
FAQ about this essay
Are open weights and open source the same thing?
No. Open weights mean the parameters can be downloaded; open source also requires a licence that permits free use, modification and redistribution. Only 23 of our 37 open-weight models are Apache-2.0 or MIT.
Why do generation models such as Wan2.2 and HunyuanVideo rank so low?
The composite covers eight dimensions and generation carries only 6%. These models score 94–95 on generation but are weak on documents, agents and Chinese, which pushes them to #51–56. For generation-only work, read the generation score and unit cost instead.
