能力 · 2026-05-12
开放权重的 AI 模型
对比在许可下公开训练权重的 AI 模型 —— 适合自托管、私域微调和受监管环境。
这是什么？
- 开放权重模型在许可（常见 Apache 2.0、MIT 或自定义开放研究许可）下发布可下载的权重。
- 注意：开放权重 ≠ 开源 —— 训练数据与代码通常并不公开。
为什么重要
- 可在自有 GPU 上自托管、用私域数据微调、离线运行，或部署到受监管环境。
- 本页价格反映最便宜的 API 托管价；若自建硬件，同一模型也可零 API 成本运行。
414 个模型支持此能力
显示前 60 / 共 414 项。 用 完整目录 进一步筛选。
Frequently asked questions
How many AI models support 开放权重?
414 canonical models in our database currently support 开放权重. The list is regenerated on every data refresh, so it always reflects the latest model releases from models.dev.
What is the cheapest model with 开放权重?
Whisper Large v3 from scaleway is currently the lowest-priced option, at $0.003 per 1M input tokens and Unknown per 1M output tokens. The full table above is sorted price-ascending.
Which model with 开放权重 has the largest context window?
Llama 4 Scout 17B Instruct (Meta) leads on context at 3.50M tokens. This may matter if you also need long-document understanding alongside 开放权重.
Which models are available on the most providers?
Production-readiness usually correlates with how many independent providers host the same weights. The top three by provider count are: Kimi K2.5 (45), MiniMax-M2.5 (40), GLM-5 (38).
How is 开放权重 different from a regular LLM?
Open-weight models publish their trained weights publicly. You can self-host on your own GPUs, fine-tune on private data or run offline. Note: open weights ≠ open source — training data and code are usually not released.
How often is this list updated?
Daily. Our data pipeline pulls models.dev once a day, regenerates the canonical model list, and rebuilds these pages so newly released models appear within 24 hours.
Explore more
Top models with this capability
- Whisper Large v3$0.00 in / $0.00 out
- Voxtral Small 24B 2507$0.00 in / $0.00 out
- KB Whisper$0.00 in / $0.00 out
- Multi-QA-mpnet-base-dot-v1$0.01 in / $0.00 out
- All-MiniLM-L6-v2$0.01 in / $0.00 out
Other capabilities
Best-of lists you might also want
Pricing comparisons
Vendors in this list
最近更新：
Prices in USD per 1M tokens. Unknown means the provider does not publish per-token pricing.
Data is sourced from models.dev and normalized for comparison. Prices and capabilities may change. Always verify critical production decisions with the provider's official documentation.