🔥剖析：小米 MiMo V2.6 Pro & Flash 不是在直播炼丹，是在给二台每小时烧掉十几万的 Agent 当场改优化方案！
This is a dev post classified by Jev as AI dev tools (news), kept by the Dev Radar because it carries real work, not commentary.
🔥剖析：小米 MiMo V2.6 Pro & Flash 不是在直播炼丹，是在给二台每小时烧掉十几万的 Agent 当场改优化方案！ 9月20日下午：Flash 停在第30步，花了 85.4万美元；Pro 跑到第28步，已烧 242万美元。两条线合计超 320万美元 / 约2200万人民币。 这三天里公开改了四刀： 1️⃣Flash 从第15步整段重跑。 基础设施错误没被及时识别，后面两步记录被覆盖。模型做错和环境挂了是两种信号，混在一起会污染评分和更新——已生成的轨迹可能整批作废。 2️⃣网络安全数据整源撤下。 轨迹日志里发现不良模式，cyber 配比归零，视觉配额上调，代码仍占约六成。 3️⃣筛掉对当前 Pro 太简单的题。 同一题试16次，全对或全错几乎没有梯度。题变难后成功率不涨，不等于模型停步。DeepSWE 上 Pro 从约64%拉到约71%。 4️⃣MoE 专家挤爆显存，Pro 触发 OOM。 改并行策略重启。每步从两个多小时拖到近四小时，条数没变，单条轨迹变长了。 结论：大规模 Agent RL 的瓶颈，已经从“有没有卡”，变成“信号清不清、题难不难、专家会不会挤爆、轨迹会不会越跑越长”。 #小米MiMo #MiMoV26 #强化学习 #AgentRL #罗福莉 #雷军
Posted by SuSu_酥酥👅 (15.6k followers) 7 h ago · 30 likes · 5.2k views · view the original post on X. Kept by the Dev Radar as AI dev tools. Tools mentioned: mimo-v2.6 RL.
More dev work like this
- 架构设计： — @mylifcc
- Multi-agent prototypes stall when orchestration becomes the project. — @DanKornas
- Rebuilding a site you own shouldn’t mean starting from a blank Next.js app. — @DanKornas
- NVIDIA wants coding agents to stop wasting tokens on repeated work. — @agenticgirl
- Runs 69 product management skills and 42 workflows in Claude Code and Cowork. — @tom_doerr
- Zai bros just opensourced their ZCode harness 🔥🔥 — @zainhas
- Your AI agents keep rediscovering work. This repo gives them a shared folder. — @DanKornas
- We made an easy image loader for Qwen Image 2.1. Now you can generate your images on the… — @Blackfrost_AI
Every post is read and classified by Jev (TypeSafe): what it is, which market it belongs to, and whether the link is a real tool. 17.6k posts from 4.9k X accounts over the last 21 days, 2k tools, 12 markets. Collected every 5 minutes, fully re-ranked every hour — last update 2026-09-21 02:20 UTC. Full method.