Open-weights ranking
Twelve models with published weights. The licence column matters more than the score — two "open" models can have wildly different commercial terms.
| # | Model | Licence | Parameters | Context | v4.3.2 | v5 | GPQA | TB 4.0 | Output $ |
|---|---|---|---|---|---|---|---|---|---|
| 1 | Kimi K3.1 Pro Open weightsMoonshot AI | Kimi K3 License | 2.8T/104B | 1M | 55.1 | 46.0 | — | — | $10.00 |
| 2 | MiMo-V2.6 Pro Open weightsXiaomi | MIT | 1.02T/42B | 1M | 46.3 | 30.0 | 59.5 | 31.3 | $1.40 |
| 3 | DeepSeek V4.1 Flash Open weightsDeepSeek | MIT | 552B/16B | 1.05M | 45.1 | 32.5 | 89.8 | 19.7 | $1.20 |
| 4 | GLM-5.3 Open weightsZ.ai | GLM-5.3 License | 753B/18B | 1M | 44.8 | 32.0 | 90.9 | 38.9 | $2.20 |
| 5 | Kimi K3 Open weightsMoonshot AI | Kimi K3 License | 2.78T/104B | 1M | 43.6 | — | 93.1 | 17.2 | $15.00 |
| 6 | GLM-5.3 Flash Open weightsZ.ai | MIT | 321B/18B | 1M | 41.8 | — | — | — | $0.80 |
| 7 | Qwen3.8 27B Open weightsAlibaba Qwen | Apache-2.0 | 27.8B | 262K | 33.7 | — | — | — | $0.40 |
| 8 | MiniMax M3 Open weightsMiniMax | MiniMax Community License | 428B/23B | 1M | 29.2 | — | 90.9 | 1.0 | $1.00 |
| 9 | DeepSeek V4 Pro Open weightsDeepSeek | MIT | 1.6T/49B | 1.05M | — | 23.0 | — | — | $3.96 |
| 10 | Gemma 4 31B Open weightsGoogle | Apache-2.0 | 31B | 262K | 19.0 | — | 75.8 | — | $0.40 |
| 11 | gpt-oss-120b Open weightsOpenAI | Apache-2.0 | 117B/5.1B | 131K | — | — | 75.8 | — | $0.20 |
| 12 | Nemotron 3 Ultra Open weightsNVIDIA | OpenMDW-1.1 | 550B/55B | 1M | — | — | — | — | $1.60 |
Open vs closed: where the gap actually is
Knowledge tasks are essentially level. On GPQA Diamond the best open model, Kimi K3, scores 93.1% against 96.4% for the best closed model, Gemini 4 Argon — a 3.3-point gap. A year ago that gap was more than 15 points.
Multi-step tool use is a whole generation behind. On Terminal-Bench 4.0 closed models reach 65.2% and open models 38.9% — 26 points. The reason is not hard to guess: agent ability depends on long-chain reinforcement learning, and closed vendors can afford that training while the open community mostly cannot reproduce it.
Licences trip people up more often than scores. MIT and Apache-2.0 are free for commercial use; the Kimi K3 licence requires a visible attribution past 100M monthly users; the MiniMax Community License and OpenMDW-1.1 each carry their own commercial limits. Read the licence before you read the score.
Self-host or call the API
When self-hosting pays
Steady high volume, data that cannot leave the network, and a need to fine-tune. That is where 30B-class models like Qwen3.8-27B or GLM-5.3 Flash hit the sweet spot.
When the API pays
Volatile traffic, a need for frontier capability, or no GPU operations team. Self-hosting a 2.8T model like Kimi K3 would take dozens of H100s — not a serious conversation.
Hybrid is the norm
Simple tasks go to a local small model, hard ones route to an API. With OpenRouter or a self-built gateway in front, cost drops below 30% of an API-only setup.