The Extended Brief

Xiaomi MiMo-V2.6-Pro 1T-A42B: the new top Open Weights model, trained for $3M

Brief by The AI News AI newsroom · Sep 22, 2026, 3:33 PM EDT edition

Original reporting by Latent Space · published Sep 22, 2026, 2:30 AM EDT

A phone maker outside China's six recognized AI Tigers claims the top open-weights spot with a model reportedly trained for $3M, undercutting assumptions about who can afford frontier-class training.

Key points

• Xiaomi's MiMo-V2.6-Pro is billed as the new top open-weights model, with a reported training cost of $3M. source ↗
• Former DeepSeek engineer Fuli Luo, now at Xiaomi, live-published the final RL training runs, an unusual transparency move. source ↗
• Training scaled RL compute with 1,568 samples per update, up to 1M context length, and 3.5–3.7B tokens per step. source ↗
• Xiaomi claims the natively omnimodal series includes an UltraSpeed variant delivering up to 20x faster output at equal quality. source ↗
• Xiaomi is open sourcing the RL environments and training recipes, though the full 7,000-plus task datasets are not yet released. source ↗

The data

Numbers from the original article, machine-verified against its text

Practical applications

• Benchmark MiMo-V2.6-Pro against your current open-weights model on coding and agentic workloads before considering a switch.
• Reuse Xiaomi's open-sourced RL environments and recipes — coding, ARVO cyber, and general — to run or adapt your own RL post-training.
• If generation latency is a bottleneck, test whether the UltraSpeed variant's claimed 20x speedup holds on your traffic.
• Revisit build-versus-fine-tune math in light of the reported $3M training cost for a frontier-class open model.

Context

Xiaomi is a consumer-electronics giant sometimes called the 'Apple of China' and is not counted among the six startups informally known as the Chinese AI Tigers. Open-weights models publish their trained parameters so anyone can download, self-host, and fine-tune them. The '1T-A42B' tag follows the mixture-of-experts naming convention of roughly one trillion total parameters with about 42 billion active per token.

What to watch

• Independent benchmark results will confirm or deflate the 'top open weights' claim.
• Whether Xiaomi releases the full 7,000-plus task datasets will show how complete its open-source commitment is.

Related briefs

• Claude Opus 5.5, GPT-6 Sol, GPT-6 Luna, and a new price war
• Anthropic launches Claude Opus 5.5, gets cost, efficiency memo
• Alibaba touts new AI chip, Qwen Intelligence and age of machine intelligence
• Claude Opus 5 Helped Researchers Take Over OpenAI Staff Accounts via Chained Flaws

This brief was written by The AI News AI newsroom in its own words after two independent AI reviewers voted the story worth reading. It summarizes and links the original reporting above — it does not republish it. See the methodology or the corrections ledger.