Xiaomi Just Made Frontier AI Free — And Silicon Valley Is Quietly Pretending It Didn't Happen
The hook nobody wants to say out loud: A phone company just beat every closed frontier lab on price-performance — and it's giving the entire recipe away.
The Controversy
Open weights just won. MiMo-V2.6-Pro scores **46** on Artificial Analysis' Intelligence Index — tying Grok 4.7, beating Grok 4.6 (44), Gemini 3.8 Flash (41), and crushing DeepSeek V4.1 (39/36). It's MIT-licensed. Download it. Fine-tune it. Own it. Pay Xiaomi **zero dollars** if you self-host.
**Meanwhile, Claude Fable 5.1 costs $60/1M tokens.** GPT-6 Astra Fast costs **$120**. MiMo-V2.6-Pro? **$1.30 total.** Flash? **$0.42** — second cheapest frontier model on Earth.
That's not a discount. That's a 100x price gap for near-frontier capability.
The Numbers That Should Terrify Closed Labs
| Model | Total $/1M | License |
|---|---|---|
| GPT-6 Astra (Fast) | $120 | Closed |
| Claude Fable 5.1 | $60 | Closed |
| GPT-5.6 Sol | $35 | Closed |
| Grok 4.7 | $8–16 | Closed |
| **MiMo-V2.6-Pro** | **$1.30** | **MIT, open** |
| **MiMo-V2.6-Flash** | **$0.42** | **Open** |
Flash trails Pro by only 2–4 points on most agent benchmarks. Same 1M context. Same multimodal. One-third the price.
How They Did It: "You Only RL Once"
- **30 massive RL steps, 750K trajectories, under 6 days**
- $2.62M (Pro) / $850K (Flash) total RL bill — pocket change by frontier standards
- **Entire agent workflows reinforced, not answers** — sequences averaging 110K–150K tokens
- Run **entirely on Chinese chips.** No Nvidia.
- Reward hacking held below 2% via "hack agents" that hunt loopholes before training
- **GRS + GAR reward systems** teach models to solve *cleanly* — smaller patches, no brittle workarounds, no swallowed exceptions
Fuli Luo, ex-DeepSeek, calls it one of the largest open-source RL runs ever: *"In an era when compute is brutally scarce, we still chose to scale up RL."*
The Real Bombshell: They Open-Sourced Everything
Not just weights. Xiaomi released 7,000+ RL task environments, the full RL framework, composable mini-harnesses, and the technical report.
This is a blueprint, not a model drop. If smaller labs reproduce Xiaomi's gains on smaller models, the "you need a trillion-dollar datacenter" moat evaporates.
What It Isn't
Not a benchmark sweep. Claude Opus 5 still wins some coding evals. GPT-5.6 Sol still leads on exploit benchmarks. Vendor-run numbers warrant skepticism. But that's not the point. A downloadable model at $0.28–$0.87/1M output tokens that's close enough doesn't need to win everything. It needs to win dollars per completed task. And it does, by a landslide.
Enterprise AI teams: the question is no longer " is open weight good enough?"
It's "why are we still paying 50–100x more for the closed stuff?"
The moat just moved from weights to workloads — and Xiaomi gave away the shovel.