2026-11-03 18:39
最后更新于2026-11-03 18:41
AI123
Mistral Large 4报告：16倍KV缓存压缩损害智能体能力
Mistral Large 4技术报告在alphaXiv平台独家发布。报告显示，将KV缓存内存激进压缩16倍会损害模型在长上下文智能体任务中的能力，即使其语言建模性能看起来没有变化。
针对这一问题，Mistral恢复了更多层的全上下文注意力，并采用DeepSeek稀疏注意力，从上下文中选取2,048个最相关的token进行关注。该模型为1T参数MoE，激活52B参数，在Terminal-Bench 4.0上得分41%，在FrontierCode 2.0上得分19.7%。其NVFP4量化检查点可适配四块GB300 GPU。
来源
- 2026-11-03 15:43 | X：RT alphaXiv: "Mistral Large 4" paper published exclusively on alphaXiv Their latest technical report shares that aggressively reducing KV-cache memory...阅读原文
RT alphaXiv "Mistral Large 4" paper published exclusively on alphaXiv Their latest technical report shares that aggressively reducing KV-cache memory can hurt an LLM's agentic capabilities, even when its language modeling performance looks unchanged. Mistral initially reduced KV-cache memory by 16x...