Xiaomi MiMo-V2.6 model training
小米直播训练 MiMo-V2.6 模型，罗福莉称沉寂半年钻研一件事
For users interested in large language models, Xiaomi's MiMo-V2.6 is a significant update that demonstrates advancements in reinforcement learning and multi-task capabilities.
Xiaomi's MiMo team has been silent for half a year after open-sourcing MiMo-v2.5 in April. They are now training MiMo-V2.6, which is in the reinforcement learning stage. The team has expanded three capabilities: computing (20B tokens per step, 1,568 prompts × 16 rollouts, fully asynchronous), environments and tools (multi-task agent reinforcement learning, mixing multiple frameworks in one run), and Grader Compute (adding compute to the scoring process with Agentic In-group Credit Assignment, including test cases and reward standards). The team will open-source these details in the coming weeks. The training cost is over US$1.25 million. A live training stream for MiMo-V2.6 is available at mimo.xiaomi.com/rl.
小米直播训练 MiMo-V2.6 模型，罗福莉称沉寂半年钻研一件事
IT之家 9 月 17 日消息，小米 MiMo 大模型负责人罗福莉今日发文，自 4 月份开源 MiMo-v2.5 后，小米沉寂了近半年，花时间研究了一件事 —— 强化学习究竟能扩展到多远 。 MiMo-V2.6 目前正处于强化学习中间阶段。该团队为其扩展了三项能力： 计算（每步约 20 亿个 Token，1,568 个 Prompt × 16 条 Rollout，完全异步） 环境和工具（多任务智能体强化学习，一次运行中混合多个框架） Grader Compute（给打分 / 评分过程本身增加算力，Agentic In-group Credit Assignment，包含测试案例和评分标准奖励） 小米 MiMo 团队将在接下来的几周内 逐步开源这些细节 。 罗福莉还直接晒出了 MiMo-V2.6 模型的训练直播链接，总花费已超 125 万美元 （IT之家注：现汇率约合 840.3 万元人民币） ，直播页面显示 MiMo-V2.6 模型即将推出 。 IT之家附直播地址： https://mimo.xiaomi.com/rl/
