MiMo-V2.6-Flash: the 309B Open Checkpoint Xiaomi Now Sells by the Token

MiMo-V2.6-Flash is the efficiency checkpoint of Xiaomi's September 2026 release: a sparse mixture-of-experts model with 309 billion parameters in total and 15 billion active per token, a 1M-token context, and MIT-licensed weights on Hugging Face. It sits below MiMo-V2.6-Pro, the 1.02-trillion-parameter flagship, inside the same MiMo-V2.6 series — and the two are separate training runs rather than one model at two sizes. What has changed since the first coverage of this release is not the model. It is everything around it. Xiaomi's own platform now publishes a per-token price against the exact identifier mimo-v2.6-flash, a published maximum output length, a rate-limit quota, and a deprecation date that will route V2.5 traffic onto this endpoint whether or not you migrate. This is the standing reference for the model itself: three published checkpoints share its name, exactly one of them is the one readers mean, and two figures you may have seen for it are now out of date.

What MiMo-V2.6-Flash is — and the two names it is not

Start with the distinction a reader has to get right, because the searchers who arrive at this page split across it. MiMo-V2.6 is a series, not a model. Xiaomi's release note says the series "comprises two native fully multimodal models, namely Pro and Flash," so the bare series name resolves to two published checkpoints with separate training runs, separate benchmark tables and separate memory footprints. Flash is the smaller, cheaper, efficiency-balanced one; Pro is the flagship. Neither is a quantised edition of the other, and there is no third "plain V2.6" sitting between them.

The second name to keep apart is older and more expensive to confuse. MiMo-V2-Flash shipped in December 2025 and, on paper, looks like the same model: a sparse mixture of experts with 309B total and 15B active parameters, hybrid attention interleaving sliding-window and global blocks, and the same 128-token sliding window. Nine months separate them and so does almost everything else. The 2025 checkpoint claims a 256K context; this one claims 1M. They are different downloads, different serving recipes and different capability profiles, and a search for "MiMo V2.6 Flash" will still happily hand you the 2025 pages.

What Xiaomi does not publish is a second weight repository under the bare name. Requesting XiaomiMiMo/MiMo-V2.6-Flash from the Hub returns an authorization error because no such repository exists; the canonical artifact is XiaomiMiMo/MiMo-V2.6-Flash-RL. More on what that suffix means below, because it is the single most misread part of this model's naming.

Who publishes it, under what licence, and under whose endpoint

Xiaomi's MiMo team publishes it, on its own Hugging Face organisation, and the licence is MIT — verified today against the repository metadata, which carries license:mit and no gating. That is the permissive licence without a revenue threshold, a research clause or an acceptable-use gate, so commercial deployment, modification and redistribution are all permitted without a conversation with Xiaomi. The model card declares English and Chinese as its languages.

The same organisation publishes three artifacts on this stem, and the distinction matters for anyone planning a deployment:

• MiMo-V2.6-Flash-RL — the canonical released checkpoint, created 21 September 2026, 90 files. This is the one the API serves and the one every benchmark in this page refers to.

• MiMo-V2.6-Flash-MOPD — published 27 September 2026, and described in its own card as "the MOPD upgrade of the MiMo-V2.6-Flash-RL checkpoint" rather than a different model. It is the output of a second post-training stage that fuses several domain-specialised teachers into one model. If you already pulled the RL checkpoint, this is a newer artifact on the same lineage, not a separate sibling.

• MiMo-V2.6-Distill-Qwen-9B — a 9.4B MIT-licensed model published the same day, which the release note names as the starting point for the open-sourced RL training run rather than a product checkpoint.

Xiaomi also sells the model directly, which is the part of this picture that moved during October. The hosted identifier is lowercase and unsuffixed — mimo-v2.6-flash — and it appears in the platform's own /v1/models listing alongside mimo-v2.6-pro, mimo-v2.6-pro-ultraspeed and the V2.5 generation.

The architecture, read from the config rather than the summary

The model card's summary line is where most secondary coverage stops, and it is not where the useful numbers are. From Xiaomi's own architecture tables and the repository's config.json, both read today:

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

• Attention heads — 64 query heads in both blocks, 8 key/value heads on the sliding-window blocks against 4 on the global ones. This is the shape that makes the 1M context affordable: the hybrid pattern is what keeps the KV cache from scaling with the full sequence.

• Encoders — a 681M-parameter MiMo ViT (28 layers, 24 sliding-window plus 4 full), a 308M audio tokenizer with 24 layers and 20 residual-vector-quantisation codebooks, and a 127M audio patch encoder that folds four frames into one.

• Speculative decoding — a 5-layer sliding-window multi-token-prediction drafter, DFlash-style, with a 1024-token window, described as predicting 7 subsequent tokens per forward pass.

• Weights on disk — the repository's file listing on the Hub today totals roughly 178 GB across 90 files, the shards in FP8 (e4m3) with a dynamic activation scheme. That is the number you pay for in VRAM, and it does not care how the parameters are counted.

Two counting questions come up here, and they now have different answers than they did in September. The first: the Hub's safetensors metadata for the repository reports a total of 310,756,322,688 parameters — about 310.8B, which lands right next to the card's 309B total. Earlier write-ups, including ours, carried a 159B figure read off the same repository panel. That number has moved, and the current one is the one to use. The second question has not been resolved: the card describes a five-layer drafter predicting seven tokens, while config.json in the same repository sets num_nextn_predict_layers to 3. Both cannot describe the same artifact, and the config is what the runtime reads. Size memory for the drafter from the config and verify after loading.

The config does corroborate one headline claim cleanly, and it is worth saying so because most of the others need a caveat: max_position_embeddings is set to 1,048,576, which is exactly the 1M-token context the card advertises.

Context, output ceiling, and a modality claim that narrows off-vendor

Context is the easy one. Xiaomi advertises 1M tokens, the config backs it, and the platform's model summary repeats it. Maximum output is the figure most reference pages still get wrong by omission, and it is published: the platform's model page gives 128K tokens maximum output against the 1M context, for both mimo-v2.6-flash and mimo-v2.6-pro. There is one behavioural trap attached. Deep thinking is enabled by default, and max_completion_tokens caps thinking content and the final answer together — so a long reasoning pass eats the same budget as the answer, and the vendor's own guidance is to set it generously. In thinking mode, custom temperature and top_p are ignored: whatever you send, the effective values are forced to 1.0 and 0.95.

Modalities are where the vendor's material and independent measurement part company, and it is worth stating both without picking a winner. Xiaomi's model card and platform both claim a native omnimodal input surface for the series — text, image, video and audio entering one model rather than three bolted-together pipelines. Independent measurement reads narrower. On the Artificial Analysis model page for MiMo-V2.6-Flash, the technical specifications list input modality as "text and image" and output as "text", while the same site's page for MiMo-V2.6-Pro lists "text, image, speech, and video" input. Same family, same vendor documentation, and only one of the two checkpoints has an independently recorded audio and video surface. Nothing published explains the difference, and we cannot settle it from outside. Treat the audio and video input as a vendor claim on Flash until something else confirms it.

What Xiaomi reports, and the one number a third party measured

Every benchmark figure Xiaomi publishes for this model comes from Xiaomi's own harness runs and has not been independently reproduced on those harnesses. That is not an accusation, it is the state of the evidence, and it is worth holding onto because the vendor's own surfaces disagree with each other in a way that is easy to miss.

• Code agent — DeepSWE v1.1: Flash 67.9, Pro 71.9, against Claude Opus 5 at 74.0 and GPT-5.6 Sol at 73.0 on Xiaomi's runs. Terminal Bench 2.1: Flash 87.6. Terminal Bench 4.0: Flash 28.8. ProgramBench 26.0.

• General agent — AutomationBench v1.0.6: Flash 52.3, above both GPT-5.6 Sol (45.8) and Claude Opus 5 (50.3) on the same vendor run. Toolathlon-Verified 73.6, OSWorld-Verified 80.8, Agents' Last Exam 27.6.

• Cybersecurity — the column where the smaller model wins. CyberGym: Flash 95.1 against Pro's 94.0. Then the floor drops: ExploitGym 6.0, ExploitBench 25.3, SEC Bench Pro 47.5.

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

The disagreement sits in the DeepSWE v1.1 column, and it runs across three Xiaomi surfaces. The release note on Xiaomi's platform says Flash improved "approximately 17 points (from 48.8 to 65.7)" across the RL run. The model card prints 67.9 for the released weights. The streamed RL dashboard that ran through September published 65.68 under a mini-swe-agent harness at average-of-three, on a training snapshot. Three vendor numbers for one benchmark on one model, and nobody outside Xiaomi can currently say which one describes the artifact you download. If you have been quoting 65.68 since September, it described a training snapshot at the time and it is now a stale number for the released checkpoint.

Against all of that there is exactly one independent measurement, and it is not flattering in the way a vendor table is. Artificial Analysis scores MiMo-V2.6-Flash at 38 on its Intelligence Index v4.3.2 — eighth of 117 models in the class — with 49.2 output tokens per second, a 240M-token generation for the index run that the site labels "very verbose" against a median of 140M, and a cost of $0.06 per Intelligence Index task. Its summary sentence is that the model is "amongst the leading models in intelligence and reasonably priced" but "slower than average and very verbose." For comparison, the same site scores MiMo-V2.6-Pro at 46 and ranks it first of 117 — the number Xiaomi itself quotes in its own release note, for Pro and not for Flash.

What it costs, and the figure that moved

This is the part of the picture that changed after the launch coverage, so it needs saying plainly: the earlier reading that Xiaomi had published no per-token rate for this generation is out of date. The vendor's pay-as-you-go page carries a full rate card with an update stamp of 8 October 2026, and the subject model has its own row on it.

• Overseas, per million tokens — mimo-v2.6-flash at $0.0028 on a cache hit, $0.14 on a cache miss, and $0.28 output. mimo-v2.6-pro at $0.0036 / $0.435 / $0.87, and mimo-v2.6-pro-ultraspeed at $0.036 / $4.35 / $8.70.

• Mainland pricing, per million tokens — flash ¥0.02 / ¥1.00 / ¥2.00, pro ¥0.025 / ¥3.00 / ¥6.00, ultraspeed ¥0.25 / ¥30.00 / ¥60.00. The platform bills in RMB for China and dollars overseas.

• Batch API — exactly half the real-time rate: flash $0.0014 / $0.07 / $0.14, pro $0.0018 / $0.2175 / $0.435, with the batch base URL separate from the real-time one. Xiaomi states that batch is not supported for mimo-v2.6-pro-ultraspeed or the V2.5 identifiers.

• Cache writes — free for a limited time, as stated on the rate card today. Cache hits are where the economics move: at $0.0028 against $0.14, a hit is 2% of a miss, and Artificial Analysis records a 98% cache discount and the same $0.14/$0.28 pair from its own provider readings.

• Web search — billed separately from tokens at $5 per 1,000 calls overseas and ¥16 per 1,000 in China, with the fetched page content then billed as ordinary input tokens. A single search round can fire several keywords concurrently, so the vendor exposes a max_keyword cap to bound the spend.

• Subscriptions — Xiaomi also sells Token Plan tiers priced per month with credit allowances and "V2.6 Flagship Model Access," a separate credential format from the pay-as-you-go key, and a night-time credit multiplier between 00:00 and 08:00 UTC+8. Subscriptions and pay-as-you-go balances are explicitly not interchangeable.

How to call it

There are three routes, and they take different credentials.

• Hosted API — the base URL is https://api.xiaomimimo.com/v1 for the OpenAI-compatible surface and https://api.xiaomimimo.com/anthropic for the Anthropic-compatible one. Authentication accepts either an api-key: $MIMO_API_KEY header or a Bearer token, and the model string is the lowercase mimo-v2.6-flash. Xiaomi's own note is blunt about the casing: use the all-lowercase names.

• Rate limits — a published 100 requests per minute and 10M tokens per minute for mimo-v2.6-flash, counted per account across all API keys rather than per key. The ultraspeed identifier is the exception, sold as a customised service rather than a metered one.

• Self-hosting — the card ships an SGLang recipe at tensor parallel 8, data parallel 2, with DP attention, the multi-layer EAGLE speculative path enabled (--speculative-num-steps 3, --speculative-eagle-topk 1, --speculative-num-draft-tokens 4), a 0.65 static memory fraction and 16384 chunked prefill, plus a vLLM recipe at tensor parallel 4. Both need --reasoning-parser mimo --tool-call-parser mimo to parse thinking output and tool calls correctly. The parallel widths in these recipes have been edited since the first write-ups of this release; use what the card says today. Recommended sampling is temperature 1.0, top_p 0.95.

One dated item belongs in any reference for this model, because it changes what an existing integration does on its own. Xiaomi's migration notice states that from 14 October 2026 at 18:00 UTC+8, requests to mimo-v2.5-pro and mimo-v2.5 will auto-route to the V2.6 series, and that from 21 October 2026 at 10:00 UTC+8 the old model identifiers stop working. Traffic pinned to a V2.5 identifier is therefore already being repointed without a code change — worth checking, because the same notice says requests to those two IDs get sent to Pro and Flash respectively, which is not a like-for-like swap on cost.

If the reason you are reading this is that you need a callable frontier model rather than a download, the honest framing is that MiMo-V2.6-Flash is not one of them here. We route no Xiaomi model, and Xiaomi's own card points availability at its first-party surfaces — the Open Platform API, its studio app, its coding tool and its desktop client. The models Xiaomi benchmarks against in its own tables are callable, though, and several of them sit on OrcaRouter on one API key at provider list price with 0% markup passed through, so a vendor price change lands on your bill the same day rather than at a renewal. That matters more than usual on a page like this one, where a rate card was published mid-flight and the totals move.

The "-rl" in the repository name, and other artifacts on the same stem

Readers search for this model as mimo-v2.6-flash-rl as often as without the suffix, so it is worth being exact. The vendor does not name an "RL variant" as a distinct model. The suffix is part of the canonical artifact's own name: XiaomiMiMo/MiMo-V2.6-Flash-RL is the released checkpoint, and -RL marks the post-training lineage — the output of the mixed reinforcement-learning run Xiaomi streamed through September, not a reinforcement-learning adapter sitting on some other base model. The supporting evidence is structural rather than rhetorical: the repository ships tens of thousands of tensors covering the whole backbone, the vision encoder, the audio stack and the speculative drafter, and the file listing is ~178 GB. An adapter is megabytes. And note what the vendor's API calls it — mimo-v2.6-flash, no suffix at all. So the "-rl" is a weights-repository convention, and typed into a hosted request it will simply not resolve.

The other three suffixes on this stem are ordinary and worth knowing: -MOPD is the further post-training upgrade published six days later, -Pro is the flagship rather than a larger Flash, and -ultraspeed is a speed mode sold as a customised service. The 9B distill and the RL task environments live on different repositories entirely. None of them is a quantisation of Flash, which is the assumption the suffixes invite.

Three questions readers arrive with

Is this the same model as MiMo-V2-Flash? No, and the confusion is expensive rather than cosmetic. The December 2025 checkpoint and this one share a headline shape — 309B total, 15B active, hybrid attention, a 128-token sliding window — but they are separate training runs nine months apart with different context claims (256K against 1M), different benchmark tables and different serving recipes. Sizing a node from the older model's figures, or quoting its price, will be wrong in both directions.

Will it fit on the hardware I have? Not on one GPU, and the vendor does not pretend otherwise. The published weight data is roughly 178 GB in FP8 before any KV cache, and Xiaomi's own recipes start at four GPUs for vLLM and eight for SGLang. Read the deployment section as the vendor's minimum rather than a suggestion, and note that the recipes need explicit reasoning and tool-call parsers — without them the model still generates, but you lose the thinking channel and structured tool calls that most of its agent benchmarks depend on.

Which Flash artifact should I actually pull? The RL checkpoint if you want what the benchmarks and the API describe, and the MOPD repository if you want the later post-training stage that its own card calls the upgrade of it. Do not pull both expecting a size difference; they are the same lineage at the same scale, six days apart, and the MOPD model is where Xiaomi's domain-teacher fusion landed.

What would change this page

Three things, in rough order of how much they would move a decision. A third-party run on the vendor's own harnesses — DeepSWE v1.1 or Terminal Bench, submitted rather than self-reported — would settle whether the 67.9 is a real position or a favourable configuration, and would also explain why three Xiaomi surfaces print three different numbers for the same column. An independent test of the audio and video input surface would resolve the gap between the card's omnimodal claim and the narrower technical note on the third-party page. And one more vendor surface reconciling the five-layer, seven-token drafter in the summary against the three layers in the config would remove the last inconsistency a capacity planner has to work around.

Until then the reference reading is narrow and usable. MiMo-V2.6-Flash is a 309B-parameter, 15B-active, MIT-licensed sparse mixture-of-experts checkpoint with a 1M-token context and a 128K output ceiling, published on 21 September 2026 as the FLASH-RL suffix, priced by Xiaomi at $0.14 per million input tokens and $0.28 per million output on a cache miss, and — as of October — callable through the vendor's own endpoint under the identifier mimo-v2.6-flash. Its vendor benchmarks are unreproduced, its only independent score is 38 on an index where its own bigger sibling scores 46, and its headline parameter count is one of the few numbers near this model that has actually come out clean.