MiMo-V2.6, Xiaomi’s new family of open-weight AI models, now has a card on AIxploria, the AI tools directory. It went up at 03:48 UTC on 28 September 2026, six and a half days after Xiaomi put the weights on Hugging Face. The card’s review is upbeat and mostly accurate. It calls the series “a rare package for developers building agents or hosting models themselves.”
The card misses the week’s biggest development. On 27 September Xiaomi published a post admitting that its new models sometimes looped, firing the same tool call again and again while making no progress. It had already swapped the fixed models into its API on 25 September under the same names, and it reset MiMo Desktop users’ quotas as an apology. None of this appears on the card, which was written a day later.
We checked the card against Xiaomi’s model cards, technical report, pricing page and repetition post, plus the independent scores from Artificial Analysis. Here is what the card gets right, what it gets wrong, and what it leaves out. For background on the wider field, see our guide to open-weight AI models.
Table of contents
- The MiMo-V2.6 Card at a Glance
- What Xiaomi Actually Released
- MiMo-V2.6 Card Claims Checked Against Xiaomi
- The MiMo-V2.6 Fix the Card Does Not Mention
- MiMo-V2.6 on the Independent Index
- Xiaomi’s Own Benchmarks, Read Carefully
- MiMo-V2.6 Pricing and What Jobs Cost
- Running the MiMo-V2.6 Weights Yourself
- The Rank Badge on the MiMo-V2.6 Card
- Where the MiMo-V2.6 Card Has Spread
- What the Card Leaves Out
- Should You Try MiMo-V2.6?
- Frequently Asked Questions About MiMo-V2.6
- References
The MiMo-V2.6 Card at a Glance
AIxploria gives every listing a blurb, a review, tags, a rating and a rank badge. The MiMo-V2.6 card follows that pattern, and several of its numbers are worth reading closely.
The basics
The card sits at aixploria.com/en/mimo-v2-6-xiaomi/. Its headline reads “MiMo-V2.6: a family of (open-source) multimodal AI models developed by Xiaomi.” It is tagged “Latest AI” and “LLM models”, marked as a Verified Tool and priced as Free. The page title promises “39 Alternatives”, but the page shows eight.
Seven votes
The card shows a 4.4 rating. The page’s structured data shows the exact figure is 4.43, from seven votes. The card also shows 71 upvotes. The Claude Opus 5.5 card, added on 24 September, also launched with seven votes.
The blurb
The short blurb describes “a series of (multimodal) LLM models able to process text and images for reasoning tasks.” Blurbs like this are how most AI tools on the directory get found, and this one undersells the product. The longer review on the same card correctly says the models also accept video and audio.
How long it took
Xiaomi’s Hugging Face repositories for MiMo-V2.6-Pro and MiMo-V2.6-Flash were created at 15:39 UTC on 21 September. OpenRouter listed both about four and a half hours later. The AIxploria card followed six days and 12 hours after the weights. AIxploria’s cards for Gemini 3.8 Flash, Claude Fable 5.1 and GPT-6 Astra came within one to three days of their launches.
| Card field | Value on 28 September |
|---|---|
| Published | 03:48 UTC, edited at 03:53 |
| Rank badge | #28 in LLM models |
| Rating | 4.43 from 7 votes |
| Upvotes | 71 |
| Price label | Free (structured data price: 0) |
| Alternatives shown | 8, of “39” in the page title |
| Useful links | Hugging Face, API, Direct Access, technical report |
What Xiaomi Actually Released
MiMo is Xiaomi’s AI model family. It began with MiMo-7B in April 2025 and moved to large mixture-of-experts models with MiMo-V2-Flash in December 2025. MiMo-V2.6 follows MiMo-V2.5, released in April 2026.
Two sets of weights and a faster API tier
Xiaomi released two sets of weights: MiMo-V2.6-Pro, with 1.02 trillion parameters of which 42 billion are active for each token, and MiMo-V2.6-Flash, with 309 billion parameters and 15 billion active. A third product, MiMo-V2.6-Pro-UltraSpeed, is an API tier. Xiaomi says it “delivers V2.6-Pro’s flagship performance at up to 20x output speed.” It has no separate weights.
What the models accept
Both models take text, images, video and audio as input and hold a context of one million tokens. They use a 681-million-parameter vision encoder and a separate audio tokeniser. Both use a speculative decoder that drafts seven tokens at a time to speed up output.
What else shipped
Xiaomi also released the technical report, its reinforcement learning framework, task environments with graders, and a 9-billion-parameter model called MiMo-V2.6-Distill-Qwen-9B. The weights carry the MIT licence, which allows commercial use. Xiaomi says it spent $2.6 million on the reinforcement learning stage for Pro and $0.9 million for Flash.
Where it runs
The models run in Xiaomi’s AI Studio, its MiMo Code and MiMo Desktop apps, its API platform and OpenRouter. The weights are on Hugging Face and ModelScope for anyone who wants to host them.
MiMo-V2.6 Card Claims Checked Against Xiaomi
We listed every claim on the card that can be checked and compared it with Xiaomi’s documents and the independent index.
| Card says | Source says | Verdict |
|---|---|---|
| Pro: 1.02T parameters, 42B active | Same (model card) | Correct |
| Flash: 309B parameters, 15B active | 309B on the model card, 310B in the report | Correct |
| Text, image, video and audio input | Same | Correct |
| 1M-token context on Pro and Flash | Same | Correct |
| Three MIT-licensed variants | Two weight sets are MIT; UltraSpeed is an API tier | Incomplete |
| Released 22 September | Weights uploaded 21 September, announced 22 September | Correct |
| Flash from $0.14, Pro $0.435, UltraSpeed $4.35 input | Same on Xiaomi’s pricing page | Correct |
| Same rates as V2.5 | Same | Correct |
| “Over 7,000 RL environments” | Report: “approximately 7k tasks” in four training sets | Rounded up |
| Artificial Analysis score of 46, top open-weight model | 46.32; next open model 44.78 | Correct |
| “Xiaomi’s own figure is 46.32” | 46.32 is Artificial Analysis’s unrounded score | Wrong attribution |
| Tied with Grok 4.7, released the same day | Grok 4.7 high: 46.33; announced 21 September | Correct |
| Ahead of Kimi K3 and Qwen3.8 Max | 43.59 and 45.42 | Correct |
| DeepSWE 72.57 for Pro, “up from 58.4” | 58.4 is Pro before RL; V2.5 Pro scored 19.0 | Needs context |
| DeepSWE 65.68 for Flash | Report chart 65.68; model card table 67.9 | Two Xiaomi figures |
| “The final training run reportedly cost $2.62 million” | The figure covers RL post-training only | Incomplete |
| 9B distill “announced for modest machines” | Released 21 September as a research starting point | Misframed |
| Spoken output via “MiMo-V2-TTS” | Xiaomi’s platform lists MiMo-V2.5-TTS | Out of date |
| OpenRouter “:free” route is the older V2-Flash | No free MiMo route listed on 28 September | Out of date |
| MiMo began with MiMo-7B in April 2025 | Same (Hugging Face) | Correct |
The score
Of 20 claims, 11 are correct as written. Two are incomplete, two are out of date, one is rounded up, one misframes a release, one misattributes a score, one needs context and one sits between two Xiaomi figures that disagree. None of the errors changes the basic picture. Several of them would matter to someone choosing a model.
Where the 46.32 came from
The review says Xiaomi’s “own figure is 46.32”. That number is the unrounded score on Artificial Analysis’s leaderboard. It is the independent figure, shown to two decimal places, not a Xiaomi claim.
What “up from 58.4” means
The review’s DeepSWE line reads as though the previous generation scored 58.4. In Xiaomi’s report, 58.4 is where MiMo-V2.6-Pro itself started before reinforcement learning. Xiaomi’s own table puts MiMo-V2.5-Pro at 19.0 on the same test, so the real generational jump is far larger than the card suggests.
The MiMo-V2.6 Fix the Card Does Not Mention
The most useful thing a buyer could learn about MiMo-V2.6 this week is missing from the card. On 27 September, Xiaomi published “Diagnosing and Mitigating Tool-Call Repetition in MiMo-V2.6.”
What went wrong
Xiaomi wrote that after launch, “tool-call repetition emerged as one of the most noticeable issues affecting the user experience.” In MiMo Desktop, MiMo Code, OpenCode and other agent tools, the model “would sometimes issue the same or highly similar tool calls repeatedly, consuming substantial time and context without making meaningful progress.”
How often it happened
Xiaomi measured one narrow kind of repetition: identical calls within a single turn. By that measure, Flash repeated itself in 1.02% of responses under OpenCode and Pro in 0.54%. Rates in other agent tools were lower. Xiaomi calls these figures a lower bound, because the measure ignores near-duplicates and repeats across turns.
| Agent tool | Flash repetition rate | Pro repetition rate |
|---|---|---|
| OpenCode | 1.02% | 0.54% |
| Claude Code | 0.27% | 0.10% |
| Codex | 0.23% | 0.07% |
| MiMo Desktop | 0.19% | 0.19% |
| OpenClaw | 0.17% | 0.08% |
| MiMo Code | 0.11% | 0.07% |
Where it came from
Xiaomi traced the problem to its training. Its reinforcement learning only penalised a turn when it made more than 32 tool calls. Behaviour below that line went unpunished and grew. In replayed test cases, the share of Flash turns with ten or more calls rose from 11.1% at the start of training to 24.6% by step 20.
The fix Xiaomi rejected
The obvious fix was a stricter limit. Xiaomi tested a cap of eight calls per turn and found it worked, but only after about 20 training steps. Redoing those steps at full scale would have cost an estimated $2.31 million. In testing, it also cut the repetition rate only from 13.45% to 3.83%.
The fix it shipped
Instead, Xiaomi trained a specialist “teacher” model on about 7,000 repetition examples over 12 steps, then folded its behaviour into the main models. The whole run cost about $90,000, which Xiaomi says is 4% of the rejected option. It reports that repetition rates “dropped substantially” while benchmark scores held.
What changed for users
The fixed models have served Xiaomi’s API since 06:00 Beijing time on 25 September, under the unchanged names mimo-v2.6-pro and mimo-v2.6-flash. The matching downloads appeared on Hugging Face on 27 September with an “MOPD” suffix. Xiaomi also reset the remaining quota for every MiMo Desktop user “as an apology.”
Why it matters
Anyone who benchmarked MiMo-V2.6 through the API before 25 September tested a different model from the one served now. Anyone who downloaded the original weights still has the looping version unless they fetch the MOPD files.
MiMo-V2.6 on the Independent Index
Artificial Analysis runs every major model through the same set of tests and publishes one Intelligence Index score. It is the most useful check on vendor claims.
Where it leads
MiMo-V2.6-Pro is the highest-scoring open-weight model on the index. The next open models are GLM-5.3 at 44.78 and Kimi K3 at 43.59. It sits level with SpaceXAI’s Grok 4.7, which we covered in our Grok 4.7 card check.
Where it trails
The top closed models are well ahead. Claude Opus 5.5 at maximum effort scores 57.62, more than 11 points higher. The card is right that Xiaomi leads the open field, not the whole field. We looked at Opus 5.5’s own card in Claude Opus 5.5 Is Now Available on AIxploria.
The jump from V2.5
Xiaomi’s previous flagship, MiMo-V2.5-Pro, scores 26 on the same index. A 20-point gain in five months is the real story of this release, and the card’s review does mention “a close look at the score jump” without giving the older figure.
Flash is a step down
MiMo-V2.6-Flash scores 37.88, about eight and a half points below Pro. On Xiaomi’s own agent tests Flash trails Pro by a few points at most, but the broader index shows a wider gap.
Xiaomi's Own Benchmarks, Read Carefully
Xiaomi’s model card compares MiMo-V2.6 with three rival models on 17 tests. The table below shows five of them.
| Test | V2.6 Pro | V2.6 Flash | V2.5 Pro | Claude Opus 5 | GPT-5.6 Sol |
|---|---|---|---|---|---|
| DeepSWE v1.1 | 71.9 | 67.9 | 19.0 | 74.0 | 73.0 |
| Terminal Bench 4.0 | 34.9 | 28.8 | 1.5 | 49.0 | 39.9 |
| OSWorld-Verified | 82.0 | 80.8 | n/a | 83.4 | 83.0 |
| AutomationBench v1.0.6 | 53.1 | 52.3 | 16.0 | 50.3 | 45.8 |
| ExploitBench | 47.9 | 25.3 | 16.6 | 70.0 | 78.5 |
The rivals are a generation old
Xiaomi compares MiMo-V2.6 with Claude Opus 5, GPT-5.6 Sol and Claude Fable 5. All three have been replaced. Claude Opus 5.5 and GPT-6 Sol both arrived on 22 September, a day after Xiaomi’s weights. The comparison was fair when written, but it is dated now.
Two DeepSWE numbers
Xiaomi publishes two DeepSWE figures for each model. The model card table gives 71.9 for Pro and 67.9 for Flash. A chart in the technical report ends at 72.57 and 65.68. The card quotes the chart. For Pro the chart figure is slightly higher than the table’s, but for Flash it is 2.2 points lower.
Security tests cut both ways
Xiaomi trained MiMo-V2.6 on cybersecurity tasks as well as coding. Flash beats Pro on one test, CyberGym, at 95.1 against 94.0. On the harder exploit tests, both trail the closed models by a wide margin. ExploitBench puts Pro at 47.9 against GPT-5.6 Sol’s 78.5.
MiMo-V2.6 Pricing and What Jobs Cost
Xiaomi’s pricing page matches the card. The rates are the same as MiMo-V2.5’s, and both V2.5 models are now marked “Offline Soon”. See our LLM API pricing guide for how these rates compare across the market.
| Model | Input per 1M tokens | Cached input | Output per 1M tokens |
|---|---|---|---|
| MiMo-V2.6-Flash | $0.14 | $0.0028 | $0.28 |
| MiMo-V2.6-Pro | $0.435 | $0.0036 | $0.87 |
| MiMo-V2.6-Pro-UltraSpeed | $4.35 | $0.036 | $8.70 |
A worked job
Take an agent job of 100 requests, each with 10,000 input tokens and 1,000 output tokens. That is one million input and 100,000 output tokens in total. On Flash it costs $0.168. On Pro it costs $0.522. On UltraSpeed it costs $5.22, ten times Pro.
The same job elsewhere
At list rates below each vendor’s long-context threshold, the same job costs $2.60 on Grok 4.7 ($2 input, $6 output), $3.00 on GPT-6 Sol ($2, $10) and $6.00 on Claude Opus 5.5 ($4, $20). Pro does it for about a fifth of Grok 4.7’s price.
Cache hits change the maths
Cached input on Flash costs $0.0028 per million tokens, 50 times less than fresh input. If 90% of the job’s input came from cache, Flash would cost about $0.045 and Pro about $0.134.
Speed is the trade-off
Artificial Analysis measures Pro at a median of 46 output tokens per second, against 83 for Grok 4.7 at high effort. That gap is what UltraSpeed exists to close, at ten times the price.
Running the MiMo-V2.6 Weights Yourself
The card’s cons list says the Pro weights are “far too heavy for desktop hardware.” Xiaomi’s own files confirm it.
How big the files are
Pro’s weights come as 132 files totalling 573.5 GB. Flash’s come as 67 files totalling 177.7 GB. Both downloads are far larger than the memory in any desktop graphics card.
What Xiaomi’s recipes assume
Xiaomi’s SGLang recipe for Pro spans two server nodes with tensor parallelism of 16. Its vLLM recipe uses tensor parallelism of 8. These are data-centre setups. Our local LLM hardware guide explains what smaller machines can realistically run.
The 9B model is for researchers
The card says a 9-billion-parameter distill was “announced for modest machines.” It was released on 21 September, and it is about 19 GB. Xiaomi describes it as a supervised fine-tune of Qwen3.5-9B, released “as a starting point for open research in agentic reinforcement learning”, not as a small chat model.
The licence
Every MiMo-V2.6 repository on Hugging Face carries the MIT licence, including the 9B distill and the MOPD checkpoints. That permits commercial use and modification.
The Rank Badge on the MiMo-V2.6 Card
The card says “#28 in LLM models.” We walked the first three of the category’s 19 pages, 36 cards in all, to see where that sits.
MiMo is not on the first three pages
MiMo-V2.6 does not appear in the first 36 positions. Slot 28 holds GPT-5.2. Page four and beyond returned a Cloudflare challenge, so we cannot say where MiMo does sit.
The list is shifted by one
We opened 18 cards on those pages and read their own badges. For slots 1 to 24, all 12 cards we checked showed their exact position. From slot 25 onward, all six cards we checked showed a badge one higher than their position.
| Slot on the category page | Card | Its own badge |
|---|---|---|
| 1 | GPT-6 Astra | #1 |
| 6 | Kimi K3 | #6 |
| 24 | GPT-5.5 Instant | #24 |
| 25 | Claude Opus 4.6 | #26 |
| 27 | Grok 4.3 | #28 |
| 28 | GPT-5.2 | #29 |
| 30 | Kimi K2.6 | #31 |
Two cards claim #28
Grok 4.3, sitting in slot 27, also carries a “#28” badge. So MiMo-V2.6 shares its number with a card from April 2026, and neither card is in slot 28.
The newest models are missing too
Claude Opus 5.5, GPT-6 Sol and Luna, and Grok 4.7, all added on 24 September, are also absent from the first 36 slots. Opus 5.5 and Grok 4.7 were both badged “#4” when they went up. The category’s first page still leads with GPT-6 Astra and older Claude and Grok models.
Where the MiMo-V2.6 Card Has Spread
A new card appears in some of AIxploria’s lists straight away and in others much later. We counted mentions on four list pages.
The new-tools lists
The “Latest AI” category’s “New AI tools” strip lists MiMo-V2.6 first and the Netlify card second. The Free AI page’s block of verified tools also shows MiMo-V2.6 first.
The ranked lists
The Top 100 page and the Ultimate List have no mention of MiMo-V2.6. That is normal for a card only hours old.
Earlier cards
The three cards from 24 September still show zero mentions on both ranked lists after four days. GPT-6 Astra, added on 6 September, had zero on both at eight days old. It now shows 20 on the Top 100 page and 10 on the Ultimate List, the same counts as four days ago. So expect MiMo-V2.6 to stay off these pages for at least a week.
What the Card Leaves Out
Besides the repetition fix, four facts a buyer would want are missing from the card.
The distillation report
Anthropic’s September threat report named Xiaomi among seven AI labs it says used Claude’s outputs to train their own models. It attributed more than 400,000 exchanges over 20 days in March and April, from more than 1,500 accounts, to Xiaomi. Several news outlets raised the report alongside this launch. We covered it in Anthropic Details Distillation Campaigns From Alibaba, Moonshot AI, and DeepSeek.
Batch pricing and plans
Xiaomi’s platform now offers a Batch API for non-urgent work, a Team plan with central billing, and a “Token Plan”. A limited offer called MiMo Claw costs ¥14.9 a month and runs on MiMo-V2.6-Pro. The card mentions none of these.
V2.5 is going offline
Xiaomi’s pricing page marks MiMo-V2.5-Pro and MiMo-V2.5 as “Offline Soon”. Teams on the older models will need to move, and the card’s statement that V2.6 costs the same makes the switch easier to plan.
MiMo-V3 is already planned
Xiaomi’s MiMo lead, Fuli Luo, has said MiMo-V3 will use a new architecture called HySparse2. Xiaomi reports that at one million tokens it cuts the first processing pass by 5.02 times and the memory cache by 4.5 times.
Should You Try MiMo-V2.6?
The card’s verdict is broadly right. MiMo-V2.6 offers near-frontier results at a very low price. The question is whether that suits your work. Compare it with the tools in our AI coding assistants guide before you commit.
Good fits
- Agent jobs that use huge numbers of tokens, where Pro’s cost per task is a small fraction of its closed rivals’.
- Teams that need to host a model themselves for data-control reasons and have data-centre hardware.
- Researchers who want a fully published reinforcement learning setup to study or copy.
Poor fits
- Work that needs the very best results, where Claude Opus 5.5 still leads by more than 11 index points.
- Security testing and exploit work, where Xiaomi’s own table shows a wide gap.
- Anyone who needs fast responses without paying for UltraSpeed.
How to test it
If you tested MiMo-V2.6 before 25 September, test again. Run the same agent tasks you already use, watch for repeated tool calls, and count the tokens each run spends. Use the MOPD weights if you host the model yourself.
Frequently Asked Questions About MiMo-V2.6
What is MiMo-V2.6?
MiMo-V2.6 is Xiaomi’s newest family of large AI models, released in September 2026. It comes as Pro and Flash weights under the MIT licence, plus a faster Pro-UltraSpeed API tier. All versions accept text, images, video and audio, with a one-million-token context.
Is MiMo-V2.6 free?
The weights are free to download under the MIT licence. The API is paid, from $0.14 per million input tokens for Flash. On 28 September OpenRouter listed no free route for any MiMo model.
What changed on 25 September?
Xiaomi replaced the API models with versions that fix repeated tool calls. The model names did not change. The matching weights were published on Hugging Face on 27 September with an MOPD suffix.
How does MiMo-V2.6 compare with Claude and GPT?
On Artificial Analysis’s index, Pro scores 46.32, level with Grok 4.7, just below GPT-6 Sol at 47.53 and well below Claude Opus 5.5 at 57.62. It is the highest-scoring open-weight model on that index.
Is the AIxploria rank badge accurate?
No. The card says #28, but MiMo-V2.6 was not in the category’s first 36 positions, slot 28 holds GPT-5.2, and another card also carries a #28 badge.
References
MiMo-V2.6-Pro-RL model card (Hugging Face)
MiMo-V2.6-Flash-RL model card (Hugging Face)
MiMo-V2.6-Pro-MOPD model card (Hugging Face)
MiMo-V2.6 technical report (PDF)
Diagnosing and Mitigating Tool-Call Repetition in MiMo-V2.6 (Xiaomi)
Xiaomi MiMo API platform and pricing
MiMo-V2.6-Distill-Qwen-9B (Hugging Face)
Artificial Analysis model leaderboard
Introducing Grok 4.7 (SpaceXAI)
Anthropic threat intelligence report, September 2026
A phone maker now has the world’s top open-weight AI model (TNW)
More AI coverage: explore Progressive Robot's AI Models, Tools & Releases hub — hands-on reviews, setup guides and benchmarks in one place.
