Best Open-Weight Models
Top open-source and open-weight AI models ranked by human preference in the Arena. MIT, Apache 2.0, and community licensed.
Data from arena.ai. Updated March 2026.
Best open model per category
| Category | Best Open Model | License | Elo | Overall Rank | Gap to #1 | Price |
|---|---|---|---|---|---|---|
| Code | GLM-5Z.ai | MIT | 1445 | #8 | 6.7% | $1/$3.20 |
| Text | GLM-5Z.ai | MIT | 1455 | #21 | 3.1% | $1/$3.20 |
| Vision | Kimi K2.5 ThinkingMoonshot | Modified MIT | 1246 | #10 | 3.4% | $0.60/$3 |
| Text-to-Image | Qwen-Image-2512Alibaba | Apache 2.0 | 1136 | #18 | 10.3% | Free |
| Text-to-Video | Kandinsky 5.0 T2V ProSberbank | MIT | 1179 | #26 | 14.6% | Free |
| Search | Diffbot Small XLDiffbot | Apache 2.0 | 1024 | #22 | 18.4% | N/A |
| Document | No open-weight models tracked | |||||
Price vs Quality
Pareto frontier shows the best models at each price point. Open-weight models in green.
Code
Proprietary leader: Claude Opus 4.6 (Elo 1548)
Text
Proprietary leader: Claude Opus 4.6 (Elo 1502)
Vision
Proprietary leader: Gemini 3 Pro (Elo 1290)
Text-to-Image
Proprietary leader: Gemini 3.1 Flash Image (Elo 1266)
Text-to-Video
Proprietary leader: Veo 3.1 Audio 1080p (Elo 1381)
Search
Proprietary leader: Claude Opus 4.6 Search (Elo 1255)
Takeaways
Code is where open-weight shines
GLM-5 (MIT) ranks #8 overall in Code — just 7% behind Claude Opus 4.6. Seven open-weight models in the top 30. For coding, the gap is closing fast.
Qwen3.5-397B is the best all-rounder
Apache 2.0 licensed, appears in top 26 across Text, Code, and Vision. At $0.39/$2.34 per million tokens, it's the best price/performance open model.
Generation is still proprietary-dominated
Text-to-Image and Text-to-Video open models lag 10-15% behind proprietary leaders. Kandinsky 5.0 (MIT) and Qwen-Image (Apache 2.0) are the best bets but won't match Veo or Gemini yet.
Document AI has zero open-weight models
The entire Document arena top 13 is proprietary. Big gap, big opportunity for open-source.