Xiaomi MiMo-V2.6-Flash Is Out: a 309B Open Model You Serve Yourself

MiMo-V2.6-Flash stopped being a training job on 21 September 2026 at 15:39 UTC, when Xiaomi's MiMo team published the checkpoint as XiaomiMiMo/MiMo-V2.6-Flash-RL on Hugging Face — ungated, MIT-licensed, technical report attached, 65 weight shards in the index. XiaomiMiMo/MiMo-V2.6-Pro-RL followed eighteen seconds later. A week before, both models were two cards marked "stopped" on a public training dashboard, and the widely repeated line about them was that no weights existed. Now they are files anyone can download and serve. That is a real change in what you can do with the model, which is a different thing from an announcement about one.

Lead with the timestamp because the release arrived without the usual vendor choreography. Xiaomi's own blog, checked on 22 September, still shows the December 2025 MiMo-V2-Flash launch as its newest post. There is no V2.6 launch essay, no published price list for the new generation, and no callable model identifier that Xiaomi has announced for either checkpoint. What exists is a repository, a technical report, deployment recipes and a set of vendor-run benchmark tables — plus a hostile little detail in the model card that only matters to people who plan to actually load the thing.

The two cards you can hold

Both checkpoints are native omnimodal with a claimed 1M-token context, and both carry the MIT licence — commercially usable, fine-tunable, redistributable, no revenue gate and no research clause. The card calls Flash the "efficiency-balanced checkpoint" of the series and Pro the flagship; the interesting part is how the two differ once you look at the configs rather than the marketing sentence.

• Parameter count — MiMo-V2.6-Flash 309B total with 15B activated per token; MiMo-V2.6-Pro 1.02T total with 42B activated. Both sparse mixture-of-experts.

• Backbone — Flash is 48 layers (39 sliding-window, 9 global attention), hidden size 4096, 256 routed experts with 8 activated, a 128-token sliding window, no shared experts. The first block is global attention with a dense feed-forward network; everything after it interleaves.

• Encoders — a 681M-parameter MiMo ViT (28 layers, 24 sliding-window plus 4 full), a 308M audio tokenizer and a 127M audio patch encoder, so text, image, video and audio all enter the same model rather than three bolted-together pipelines.

• Speculative decoding — a five-layer multi-token-prediction drafter, DFlash-style, described in the card as predicting seven tokens ahead per forward pass for parallel verification.

• Storage — the published index reports 172.9 GB of weight data across 65 shards, in FP8 (e4m3) with a dynamic activation scheme. Read that as roughly nine bytes per parameter: Xiaomi shipped the big model, not a laptop-sized quantisation of it.

Three numbers that do not agree

This is worth stating plainly, because all three affect a deployment decision rather than a benchmark argument, and none of them is sinister — they look like documentation drift between the write-up, the repository page and the shipped files.

The first is the speculative decoder. The model summary says the MTP head is a five-layer drafter predicting seven tokens per pass. The config.json in the same repository sets num_nextn_predict_layers to 3. Both numbers cannot describe the same artifact, and the config is the one the runtime will read. If you are sizing memory for speculative decoding, trust the config and verify after loading.

The second is subtler and is not an error at all so much as a naming convention that reads like one. The repository is called MiMo-V2.6-Flash-RL, which invites the assumption that it is a reinforcement-learning adapter rather than a model. It is the model. The -RL suffix marks the post-training lineage — this checkpoint is the output of the mixed RL run Xiaomi streamed, not a LoRA sitting on top of some other base. The weight index confirms it: tens of thousands of tensors covering the full backbone, the vision encoder, the audio stack and the drafter.

The third is the one you meet first if you size hardware from the repository page instead of the card. The Safetensors block on that page reports 159B parameters. That is neither of the card's two numbers — not the 309B total, not the 15B active — and it is the figure a capacity planner is most likely to copy, because it sits next to the download button. Xiaomi has not explained the difference and we cannot settle it from outside; the likeliest reading is a counting convention over quantised tensors rather than a different model. The safe move is to size from the published weight data instead: 172.9 GB of FP8 across 65 shards is a number that has to be paid for in VRAM regardless of how the parameters are counted.

What the benchmarks say, and who ran them

Every figure below comes from Xiaomi's own evaluation tables in the model card. None of it has been independently reproduced, none of it appears on a public leaderboard, and the comparison columns are the vendor's own runs of the same harnesses against other companies' models. Treat the shape of the results as informative and the decimal places as decoration.

• Code agent — DeepSWE v1.1: Flash 67.9, Pro 71.9, against Claude Opus 5 at 74.0, GPT-5.6 Sol at 73.0 and Claude Fable 5 at 70.0. On Terminal Bench 2.1, Flash lands at 87.6 against Opus 5's 89.1 — a gap small enough that the harness, not the model, may be deciding it.

• General agent — AutomationBench v1.0.6 puts Flash at 52.3, above both GPT-5.6 Sol (45.8) and Claude Opus 5 (50.3) on Xiaomi's run. Toolathlon-Verified: Flash 73.6, Pro 76.9, Opus 5 80.6.

• Cybersecurity — the most lopsided column in the table. CyberGym: Flash 95.1, above its own bigger sibling at 94.0, with MiMo-V2.5-Pro at 40.0. Then the floor drops out: ExploitGym 6.0 for Flash against 22.1 for Opus 5, ExploitBench 25.3 against 70.0, SEC Bench Pro 47.5 against 79.1 for GPT-5.6 Sol.

• Visual agent — MiMo VisualCoding: Flash 71.5, Pro 72.3, Opus 5 70.0.

One number is worth pulling out because it is the kind of thing a launch post usually hides. Xiaomi's RL dashboard published Flash at 65.68 on DeepSWE v1.1 — and the model card now prints 67.9 for the same benchmark on the finished checkpoint. Those are not the same measurement. The dashboard figure was labelled mini-swe-agent, avg@3, on a training snapshot; the card's table is the vendor's offline evaluation of the released weights. The two-point difference is the gain from the last part of the run plus whatever changed in the evaluation setup, and nobody outside Xiaomi can currently separate the two. If you have been quoting 65.68 since last week, it is now a stale number for a different artifact.

What it costs to actually run it

Open weights are a licence, not a bill, and this is where the release gets less celebratory. The card's own deployment section recommends SGLang (tensor parallel 16, data parallel 2, with the EAGLE-style multi-layer speculative path enabled) or a vLLM recipe at tensor parallel 8, and notes that stable vLLM may lag the architecture. That is a multi-node serving job for a 309B model whose weights occupy about 173 GB before you add KV cache for a 1M-token context. Recommended sampling is temperature 1.0, top_p 0.95.

So the honest framing of "open source" here is: Xiaomi removed the licence barrier and left the hardware barrier exactly where it was. Fine-tuning Flash on your own traces is now legal and possible; doing it on anything smaller than a serious GPU node is not. That is a real and different offer from a hosted endpoint, and it is why the two ways of using this generation do not compete — they cover different budgets.

If you want the capability without the node, the models Xiaomi benchmarks against in that table are the practical alternative: GPT-5.6 Sol, Claude Opus 5 and Claude Fable 5 are all callable today, and they are one API key at provider list price with 0% markup passed through on OrcaRouter, so the decision you are actually making is "buy a node" versus "meter the calls". If you want to see how the callable tier compares on the same coding tasks before committing either way, the coding board is a faster read than the vendor table. What you cannot do on any router today is call MiMo-V2.6-Flash itself — we route no Xiaomi model, and the availability line in Xiaomi's own card points at Xiaomi's own channels: the MiMo API platform, AI Studio, MiMo Code and the desktop app.

The pricing question is still open

Xiaomi has not published a per-token rate for MiMo-V2.6-Flash. Reporting on the launch says the series will retain MiMo-V2.5's API pricing, and that is a press claim rather than a vendor price sheet — the V2.5 generation's published overseas rates were around a tenth of a dollar per million input tokens with cached input an order of magnitude cheaper, but nothing on Xiaomi's site confirms the new checkpoints inherit them. Until a price list appears, any figure you see for a MiMo-V2.6 endpoint is someone's inference.

Two other things are still missing, and both are on Xiaomi's own to-do list rather than anyone else's. Luo Fuli's statement during the RL livestream was that the training environments and the RL code would be open-sourced, and the launch material repeats that commitment; the repositories currently ship weights, a technical report and deployment instructions, not the training stack. And there is no independent evaluation: no leaderboard submission, no third-party rerun of the DeepSWE or CyberGym columns. That is the gap that matters most for anyone deciding whether a 309B model with a 15B active budget is worth a node.

What would change the picture

Three signals are worth watching, in rough order of how much they would move a decision. A published price list turns this from a self-hosting project into a line item you can compare against the hosted frontier. A third-party run on the same harnesses — DeepSWE or Terminal Bench, submitted rather than self-reported — would settle whether the 67.9 is a real position or a favourable configuration. And the RL environments, if they land, would be the more interesting artifact for most teams, because they are what you would need to reproduce the self-improvement loop on your own data.

Until then, the practical reading of this release is narrow and clear. MiMo-V2.6-Flash is a legitimate open-weight, MIT-licensed, omnimodal agent model at a size that assumes a cluster — and it is the first checkpoint in the MiMo-V2.6 generation you can hold in your hands, which is more than could be said for it last week.