Xiaomi published the middle of the recipe

Most labs hand you a finished model and a benchmark table. On Monday, Xiaomi’s MiMo team did something more interesting: they released the middle of their process. Alongside two finished checkpoints, they published a 9B model whose model card says, in as many words, that it exists “as a starting point for open research in agentic reinforcement learning” — plus the technical report, the full training-data mixture, and on GitHub, a repo called MiMo-Code (13,267 stars) whose tagline is “where models and agents co-evolve.”

The bet behind that is worth understanding, because it’s a bet on the open ecosystem. The capability gains right now are in RL on agentic tasks, and agentic RL doesn’t mostly need compute — it needs environments, harnesses, and people who enjoy breaking things. That is exactly what the open world has and what no single lab can scale alone. So they gave away the starting line and the recipe, and invited everyone else to run the leg that multiplies.

The self-improvement loop, in plain terms. RL trains models by rewarding what works, but pass/fail is a terrible teacher: if every rollout passes — or all of them fail — there is nothing to learn from. MiMo V2.6’s answer is a grader that reads the successful solutions and ranks them against each other: it builds rubrics from contrasting rollouts, then shifts training advantage toward the better ones. Two details in that design matter if you run agents. First, the loop explicitly rewards shorter paths and fewer tokens per task — that is not an efficiency footnote, it is your token bill and your agent’s patience. Second, they trained one mixed run across coding, agents, vision and security instead of separate specialists, and claim the strategies transfer to harnesses the model never saw during training. If that survives independent testing, it is the difference between a model that aces its training rig and one that survives your rig.

The part you can run today is the 9B. MiMo-V2.6-Distill-Qwen-9B is a Qwen3.5-9B fine-tuned on 77.4 billion tokens of Xiaomi’s own agent data — and the card obeys the one rule that matters most: it prints the base model on the same rows, same harness, vendor-run.

That is a +39% jump on repo-level coding, a sixfold gain on office automation, and a leap on security work — on a 9B that runs on anything you own (18.8 GB at full precision), and it carries vision: it is an image+text model, not text-only. Day-one MLX builds already exist in 4-bit, 6-bit and 8-bit. One caveat before you get excited: several of those builds are tagged text-generation, not vision — “image-text-to-text” on the card is a claim about the model, not about your serving engine. Whether MLX fully serves MiMo’s vision path yet is exactly the kind of question worth answering before calling it runnable.

The big one, for scale. Flash-RL is the efficiency checkpoint of the family (the Pro model is the flagship): about 173 GB of weights as shipped, a sparse mixture-of-experts (8 experts fire per token), a 1M-token context window, native text/image/video/audio, MIT license. On a 512 GB Mac it fits with plenty of KV headroom even unquantized; on a Spark it is a quantization job whose good recipes don’t exist yet. Their own serving recipe is SGLang with built-in speculative decoding — datacenter-flavored. The local recipes will come from us.

It’s harness week, apparently. Two more arrived while MiMo was still warm, and together they argue the same thing from the outside: the model is only half the agent — the other half is the scaffolding, and the scaffolding has opinions.

StepFun shipped Step Code, a terminal coding agent in the Claude Code mold. Two ideas worth stealing even if you never sign in: long-horizon tasks get split across parallel subagents with isolated contexts, so the main conversation stays clean while work runs for hours; and /goal hands an objective to the agent and lets it drive autonomously, live timer running, with /cron for scheduled work. It is tuned alongside StepFun’s models for token efficiency, and most Claude Code plugins work as-is. The catch for local-weights people: it’s Step-provider only — your models don’t get invited. Worth watching anyway, because community mirrors of a Step-5-Preview just surfaced on Hugging Face, and if official weights land, the first question is the one MiMo just taught us to ask — not “how fast,” but “what actually runs.”

Unreal Agent went further and put numbers on it. A Sequoia-backed team of ex-DeepMind, CERN, Meta and Bloomberg engineers built a harness whose thesis is that agents waste tokens managing their own tools — the waits, the polls, the heartbeats. Their fix: fully asynchronous tool calls. The moment a tool is issued it runs in the background while the agent keeps thinking, and results land in the session log when they finish — appended in a way that doesn’t break the KV cache, which is exactly the kind of unglamorous detail that separates engineering from slideware. Running GPT-6 Astra xhigh, they match Codex’s pass rates on Terminal-Bench 4.0 (57.9% both, $1,428 against $2,350) and SWE-Atlas (65.8% vs 63.3%, $936 against $1,303) while spending 16–39% less — the “up to 40%” headline is the best case, and the tables show the range. They publish the Harbor run IDs, so you can check their homework. Go SDK, open on GitHub, and it arrives alongside a paper asking the right question outright: HarnessTax — how much does the harness matter for coding agents?

Verdict: nothing here is worth the setup trouble yet — and that’s fine. It’s day one: the community quants have zero downloads because they are hours old, nobody independent has measured anything, and the leading-edge move is a watchlist with names on it, not an evening of yak-shaving.

1. 4-bit vs 6-bit fidelity on the 9B — the moment someone publishes quality-vs-base numbers instead of just files.
2. The 9B on a real harness against your current agent worker — the cheapest genuinely-new worker you could trial.
3. A quant recipe for Flash-RL on 128 GB+ machines.
4. Whether the intermediate checkpoints keep coming — that’s the public-training bet, paying off or not.
5. Whether local vision support for MiMo lands in the engines — the model has eyes; the question is whether your runtime does.
6. Official Step-5 weights — the mirrors exist; the release doesn’t yet.

Because that is the real story. If “release the middle of the process” works, labs get a world of free environments and eval diversity, and we get agents trained on harnesses like ours. If it doesn’t, we still got one good 9B out of it — which, at 18 GB, is already nothing to sneer at.