Moonshot AI 发布 2.8T 参数 Kimi K3.1 Pro 权重
月之暗面把 2.8T 参数的 Kimi K3.1 Pro 权重直接开源了，4-bit 发布省存储，编程基准分数贴近 Claude Opus 5.5，纯知识题却差 GPT-6 一截，许可证对云托管商还有限制。
Moonshot AI 在 Huggingface 开放 2.8T 参数的 Kimi K3.1 Pro 权重，采用 4-bit 训练并按同精度发布，避免了 16-bit 副本所需的 5.6TB 存储空间。模型每次推理仅激活 104B 参数，93 层中有 69 层使用线性注意力，使 1M-token 上下文的处理成本更低。Moonshot 官方测试显示其在 Terminal-Bench 4.0 上得分 65.7%，接近 Claude Opus 5.5 的 66.4%，但在无工具情况下 HLE-Diamond 仅得 39.4%，低于 GPT-6 Astra 的 65.8%。开源前团队还从预训练语料和 RL 环境中移除了引发网络安全攻击能力上升的内容。许可证采用修改版 MIT，年收入超 2000 万美元的托管服务商需另行协商授权。
Moonshot AI released the massive 2.8T-parameter Kimi K3.1 Pro weights on Huggingface.
On Terminal-Bench, its head-to-head with Claude 5.5
> Trained with 4-bit weights, it also ships at that precision, avoiding the 5.6TB a 16-bit copy would need.
> Moonshot traced steadily rising exploit skill to cyber content buried in its general mid-training corpus and RL environments, and removed both before opening the weights.
> In Moonshot’s own tests the model scores 65.7% on Terminal-Bench 4.0, near Claude Opus 5.5’s 66.4%, but without tools reaches only 39.4% on HLE-Diamond, against GPT-6 Astra’s 65.8%.
> Only 104B of its parameters fire per token, and 69 of its 93 layers use linear attention, which makes a 1M-token context cheaper to process.
> Its modified MIT license exempts internal use but makes model-hosting firms earning over $20M a year negotiate, reining in resellers that could undercut the company’s own API.
