Xiaomi's MiMo-V2.6 Tops Open-Source Rankings After $3.5M Training Spend
nashnova research
Xiaomi spent six days and $3.5 million on reinforcement-learning training for MiMo-V2.6, making it the top-scoring open-source model globally — the first time an open-source model matches frontier closed-source intelligence at a fraction of the per-task cost.
$3.5 million in six days — what does that price tag mean?
Xiaomi spent roughly $3.5 million over six days to train the Pro version via reinforcement learning; the lighter Flash version cost about $900,000 in roughly three and a half days.
This means → large-scale RL training now has its first public price tag: no Chinese AI lab had previously disclosed an explicit cost breakdown for this process.
In plain terms = "How much does it cost to train a frontier model?" used to be rumor. Xiaomi just laid the receipt on the table — $3.5 million buys global open-source No. 1.
What does open-source No. 1 actually look like?
Pro carries 1.02 trillion parameters with 42 billion active, scoring 46 on the Artificial Analysis Intelligence Index — the highest among all open-source models worldwide.
Flash, at one-quarter the cost, showed a larger relative gain: after 30 training steps its average pass rate rose 25%, versus 12% for Pro.
On the out-of-sample coding benchmark DeepSWE v1.1 — a test of an AI agent's ability to work continuously in real codebases — Pro climbed from 58.4 to 72.57 and Flash from 48.7 to 65.68. Both OpenAI and Anthropic now list this benchmark as a key evaluation for frontier releases.
Same 46-point score — how much cheaper is open-source?
MiMo-V2.6-Pro completes one Intelligence Index task at an average cost of just $0.13, the first time a 45-plus frontier model enters the sub-$0.10 range per task.
Closed-source Grok 4.7, released the same day, also scored 46 — but its xHigh tier averages $3.74 per task, roughly 29 times more expensive.
This means → at equivalent intelligence, the cost gap between open-source and closed-source has widened to one to two orders of magnitude. Xiaomi says Pro runs at 1/20 to 1/60 the cost of leading international models.
What exactly did Xiaomi open-source?
The release includes Pro and Flash model weights, a full technical report, over 7,000 RL task environments, an end-to-end RL training framework, and modular mini-harnesses.
A distilled variant — MiMo-V2.6-Distill-Qwen-9B, fine-tuned on Qwen3.5-9B with 77.4 billion tokens of MiMo-generated data — is also open.
In plain terms = Xiaomi didn't just hand out the model; they published "how we trained it" and "what problems we trained on" — enough for other teams to reproduce the process.
From coding to scientific research — where is the capability boundary?
Internally, Xiaomi applied Pro to designing novel metal–organic frameworks (MOFs) — a class of materials aimed at adsorbing per- and polyfluoroalkyl substances (PFAS), a frontier challenge in materials science.
With no domain-specific training, the model independently conducted literature search, hypothesis generation, novelty checking, and molecular simulation, producing two candidate structures whose simulated PFAS adsorption was a million to ten million times that of control materials.
Professor Dou Jinhu of Peking University called Pro's performance on the project "equivalent to a well-trained doctoral researcher." Xiaomi says the full R&D cycle shrank from one month to two or three days.
What comes next?
MiMo lead Luo Fuli said the training "exceeded DeepSeek-R1 in both research innovation and engineering challenge" — a project she herself had worked on. Hugging Face's CEO called it "amazing."
Ai2 post-training lead Nathan Lambert praised the combination of "top-scoring open-source model plus a public RL dashboard," adding that "anyone who truly understands open-source models already knew Xiaomi was building something big with MiMo."
This reflects a competitive shift: the open-vs-closed race has moved from "who is smarter" to "who is cheaper at the same intelligence" — whether that cost gap keeps widening is the core variable for the long-term viability of the open-source route.
市场有风险，内容仅供研究参考，不构成投资建议。