Updated May 24, 2026
llmleaderboard.in
Best Open Source LLM in 2026
Models with open weights or open licensing — self-host, fine-tune, or run via low-cost APIs. Ranked by benchmark scores where available.
Llama 4 Maverick and Qwen 3.5 397B lead open-weights models on reasoning benchmarks. Llama 4 Scout adds a 10M context window and very high inference speed when hosted on supported hardware.
| # | Model | Provider | GPQA | SWE-Bench | Context |
|---|---|---|---|---|---|
| 1 | Qwen 3.5 397B | Alibaba | 84.2% | 72.1% | 128K |
| 2 | Llama 4 Maverick | Meta | 80.1% | 62.3% | 1M |
| 3 | Qwen 3 235B | Alibaba | 76.8% | 60.5% | 128K |
| 4 | Llama 4 Scout | Meta | 76.5% | 55.8% | 10M |
| 5 | Sarvam 105B | Sarvam AI | 66.5% | — | 128K |
| 6 | Sarvam 30B | Sarvam AI | — | — | 32K |
Self-hosting vs API
Open weights reduce vendor lock-in and can lower cost at scale, but you pay for GPUs and ops. Many teams use open models via hosted APIs (Together, Fireworks, Groq) for speed without managing infrastructure.
See all 45 models with live benchmarks, speed, and pricing.
Open full LLM leaderboard →