Yesterday’s decision-model dispatch was about a typed if that refuses to write a sentence. This morning a phone company dropped MIT weights that write, see, hear, and run agents — and they sit at the top of the open Artificial Analysis board.

Xiaomi’s MiMo-V2.6 series is live. Pro, Flash, and a 20x UltraSpeed serving mode. Weights on Hugging Face. Same API prices as V2.5. The closed labs still win the absolute index. The open-weight race just changed hands.

A six-day RL run, not a new pretrain story

Xiaomi’s launch writeup frames V2.6 as an RSI step: scale reinforcement learning on verifiable, messy tasks until the model keeps finding the next capability on its own. They kept the V2.5 prices. They spent the compute on the post-train.

The closed comparison is the honest one. Xiaomi says Fable 5.1 and GPT-6 Astra both sit around 53 on the same index. Pro beat Kimi K3 (44) and GLM-5.3 (45) and jumped 20 points from MiMo-V2.5-Pro’s 26. That is the open-weight crown, not the frontier.

The RL bill is the part worth staring at. Pro and Flash each ran 30 steps in under six days, about 750,000 trajectories, at reported costs of 2.62 million and 850,000 dollars. Average pass rate on the training mix rose 12% and 25% relative. On held-out DeepSWE v1.1, Flash went 48.8 → 65.68 and Pro 58.4 → 72.57. Xiaomi’s three knobs: bigger batches (1,568 samples, 1M-context training, 3.5–3.7B tokens per step), more mixed environments (code, general agents, visual, cyber in one run), and more grader compute so long-horizon rewards are not noise.

You Only RL Once is the slogan on the Flash-RL card. One mixed batch, multiple harnesses, so a strategy learned in a SWE sandbox transfers to a harness the model never saw. That is the same lesson as yesterday’s EnvHarness, pointed at the policy instead of the wrapper.

They also claim a Lean 4 formalization of Li–Yorke’s “period three implies chaos” — 6,000-plus lines, kernel-checked, no Lean-specific post-train. Treat that as a demo, not a Fields Medal. The useful artifact is the training recipe, not the theorem.

The 9B you can actually run

Nobody is loading a trillion-param MoE on a consumer GPU tonight. Xiaomi knows that. Alongside Pro and Flash they released MiMo-V2.6-Distill-Qwen-9B: an SFT of Qwen3.5-9B on MiMo-generated trajectories, plus 7,000-plus RL task environments covering software engineering, vuln reproduction, knowledge work, and visual coding.

GGUF and MLX quants showed up inside a day. That is the local research seed. The big checkpoints are MIT, which is the rare honest license this month — not a research-only leash — but Xiaomi still publishes no hardware guidance for self-hosting Pro or Flash. A community 4-bit MLX conversion of Flash is already measuring ~164 GB peak unified memory on a 256 GB Mac Studio. Plan accordingly.

Serving is the other half. MiMo Desktop left early access. Pro-UltraSpeed is billed at up to 20x output speed at the same quality. API, OpenRouter, AI Studio, MiMo Code. If you just want tokens, you do not have to own the MoE.

Hangzhou answers with a chip and a 10-trillion roadmap

While Xiaomi was pushing weights, Alibaba was on stage at Apsara. The official release is the full-stack version of the same RSI pitch, except they own the fabless chip line too.

Qwen 4 is in training. Qwen 4.5 and Qwen 5 are mapped to 5 to 10 trillion parameters, two to four times Qwen 3.8 Max’s 2.4T. Eddie Wu’s other number: machine thinking is still under 3% of human thinking, and he wants 1,000x. That is a slide, not a ship date.

The empirical RSI claim is more checkable. Alibaba says a fully automated month of pipeline design, data validation, and error diagnosis ran 33 cycles on Qwen3.8-Max and moved its Artificial Analysis score from 40 to 45. In a chip-design loop the model made more than 10,000 EDA tool calls over 60 hours and cut bus-module area 42% with no performance loss — vendor eval, file under “interesting if independent.”

The silicon: T-Head’s Zhenwu V900, called China’s most powerful AI chip on stage. Triple the prior M890 (May). 216 GB memory, 1,200 GB/s inter-chip, native FP8/FP4. Mass production and commercial release Q1 2027. Clusters up to 500,000 cards. Alibaba Cloud wants more than 20 GW of global datacenter capacity by 2032. Shares jumped about 5% on the day.

Qwen-Image 3.1 is promised later this year with native transparent backgrounds. Yesterday’s Qwen-Image-2.1 was the 7B research-license compositor. Do not mix the two. One is weights you can download. The other is a keynote.

If you are buying a box to run the 9B distill, the affiliate math is the same as always: a used RTX 3090-class card still clears it, and Sutton and Barto’s Reinforcement Learning is still the book people pretend they finished.

OpenAI will share the proofs. It will not share the brake.

On the closed side, OpenAI spent Monday standing up an Advisory Group on Mathematics and Artificial Intelligence at the Institute for Advanced Study. The pretext is a new internal model, training started August 28, that OpenAI says resolved more than 100 long-standing open problems across most of mathematics — on top of last week’s Navier–Stokes claim. The list is not public. Independent checking is not public. Hodge is in the rumor mill. File the count as a press number until a proof index exists.

The group is the actual product. Initial names: Timothy Gowers, Edward Witten, Martin Hairer, Ravi Vakil, Melanie Matchett Wood, and four others. Unpaid. Free to publish advice OpenAI did not ask for. Free to criticize. Explicitly not responsible for advising OpenAI on how fast to pace internal math research. They get a say in how results are reviewed and released. They do not get a say in whether the next thousand problems get solved before anyone has read the first hundred.

That is a direct answer to the Fields-heavy open letter A Severe Misalignment of AI in Mathematics: stop treating “number of open problems closed” as the scoreboard, because the scoreboard eats the conceptual culture. Gowers joined the group and did not sign the letter. His version of the risk is blunter — people who would have become custodians of the tradition just pick a different dream.

Useful governance, then, or a communications layer around a firehose. Either way it is not a slowdown.

What to do with this

If you run agents, the download is the 9B distill plus the 7k environments, not the 1.02T MoE. If you buy tokens, Pro at V2.5 prices is the open-weight workhorse until someone independent reruns DeepSWE. If you watch capital, Alibaba just told the room the next Qwen is a cluster-scale object and the chip to train it ships in Q1. If you watch labs, OpenAI will keep minting theorems and an IAS committee will argue about the press release.

The closed index is still 53. The open index is now a phone company.

Want this in your inbox every morning? Subscribe to the SpaghettiStories newsletter.

Some links may be affiliate links. If you’re buying hardware to run local models, this affiliate link helps keep the lights on.