Xiaomi releases MiMo-V2.6 open-weights models, hires ex-DeepSeek engineer who ran training in public
2026-09-21
Xiaomi released MiMo-V2.6-Pro and MiMo-V2.6-Flash, described in Latent Space's Sept 22, 2026 AI News roundup as natively omnimodal open-weights models, plus a MiMo-V2.6-Pro-UltraSpeed variant claimed to deliver up to 20x faster output at the same quality. The roundup credits Fuli Luo, described as a former DeepSeek engineer now at Xiaomi, with publishing RL training metrics live during final training runs, and says Xiaomi is open-sourcing training environments and recipes (coding, cyber, general agent, visual/web-dev, music) but not yet the full 7,000+ task datasets. The roundup cites a claim, attributed to the Artificial Analysis account on social media, calling it 'a new top Open Weights model, trained for $3M' — unverified beyond that repost.
Significance 3: A potentially major new open-weights entrant and a documented talent move from DeepSeek, but the capability and cost claims are lab-and-aggregator-relayed via a social media roundup with no primary source or independent benchmark, holding this below a verdict-changing 4 or 5 per the facts-vs-opinion rule.
Implications · machine-drafted, not owner judgment
If Xiaomi's claims hold up, it becomes another credible Chinese open-weights lab (alongside DeepSeek, Qwen, Z.ai/GLM, Moonshot, Tencent) racing at the open frontier, further narrowing the gap the open-weights-one-generation-behind thesis tracks. Fuli Luo's move from DeepSeek is a data point that senior open-model talent is diffusing across Chinese labs rather than concentrating, which could accelerate the field's overall cadence.
- An independent benchmark (Artificial Analysis, LMArena) placing MiMo-V2.6-Pro against DeepSeek V4 or GLM-5.3
- Confirmation of the reported ~$3M training cost from a source other than social media
Sources
- curator [AINews] Xiaomi MiMo-V2.6-Pro 1T-A42B: the new top Open Weights model, trained for $3M retrieved 2026-09-22