Open-Weight AI Models
21 open-weights models ranked by preference Elo and SWE-bench. Read our open vs closed evaluation guide.
Open-weight-adjacent DeepSeek flagship. High reasoning density per dollar.
Alibaba specialized open reasoning model competing with frontier closed reasoning models.
Zhipu flagship. Strong Chinese/English coding and agents.
DeepSeek Jul 31 2026 price-performance SKU. Public listings put it near $0.14/$0.28 per 1M tokens.
MiniMax coding model. Tied near the top of official SWE-bench bash-only in Feb 2026.
Zhipu high-throughput multimodal vision model with sub-second latency and competitive coding.
Meta natively multimodal open-weight flagship.
Alibaba dedicated open-weight code generation model with near-frontier SWE-bench Verified coding capability.
Meta open-weight Llama 4 long-context sibling of Maverick. Common public API lists sit near $0.08–$0.30 / $0.30–$0.70 per 1M; we store a conservative hosted list until OpenRouter overwrites.
Alibaba open-weight 27B drop dated Aug 14 2026. Dense enough to self-host; not a frontier MoE.
Meta flagship open-weight 405B dense foundation model with 128k context window.
DeepSeek open-weight Mixture-of-Experts coding model supporting 338 programming languages and 128k context.
Open-weights reasoning model trained with large-scale RL.
Llama 3.1 70B instruct. Predecessor to 3.3 70B and Llama 4.
Open-weights non-commercial base model with 12B parameters, serving as the foundation for the open-source fine-tuning ecosystem.
Apache 2.0 4-step distilled open weights model optimized for local inference and instant preview generation.
Open-weight release of Ideogram 4.0, bringing world-class text rendering to self-hosted enterprise infrastructure.
Alibaba high-efficiency open video/image unified transformer backbone.