TL;DR: MiMo-V2.6-Pro is Xiaomi's MIT-licensed open-weights flagship, sold at $0.435 in and $0.87 out per million tokens, and it scores 30 on the Artificial Analysis Intelligence Index v5.0.1 at $0.26 per index task. In our seven tests in Xiaomi MiMo Studio on November 20, 2026, it passed three cleanly, half-passed three and declined one, about 4.5 out of 7. Use it through the API for cheap coding and structured-output jobs, not as a chat assistant.
MiMo-V2.6-Pro spent between 400 and 494 seconds writing a password checker in Xiaomi's own chat app, then solved a word problem in 4 seconds. Same model, same session; the gap is how long it decides to think.
Artificial Analysis paid about a fourteenth as much per task to run it as it paid for Kimi K3.1 Pro, the open-weights leader.
What we found
1MiMo Studio, the free way to try it, wants a Xiaomi account and calls itself a developer demo platform.
2It found the cause of our sort bug in 1 min 43 s, then got backwards which customers the bug hits.
3Its checker rated "Password123!" Medium, 54 of 100, where ChatGPT, DeepSeek, Qwen and GLM called it strong on November 17.
4Asked about a paper that does not exist, it refused to invent findings but said its "searches" came up empty. Studio has no search.
Verdict: Is MiMo-V2.6-Pro Worth Using?
Yes, for batch and agent work billed by the token; no, as a chat assistant. On short, checkable jobs it was quick and right: 276 on the math prompt in 4 seconds, clean JSON in 2 seconds.
On long code it was slow but careful, and its password checker graded harder than most tools we have given the same prompt. I would put it behind an API for batch jobs and keep a faster model for chat.
The misses were all about trust. It reversed one step in a correct debugging answer, dropped a period from a line we asked it to copy exactly, and claimed a search it could not have run.
None of these broke a deliverable, but each one is the kind of slip a reviewer has to catch.
Strong in our tests
Arithmetic: 276, right, in 4 s
JSON only: valid, no fences, in 2 s
Code: a working checker with strict grading
Weak in our tests
Current facts: declined, no web search
Exact wording: dropped a required period
Honesty about tools: claimed a search it cannot run
What MiMo-V2.6-Pro Is
MiMo-V2.6-Pro is Xiaomi's largest open model: a 1.02-trillion-parameter mixture-of-experts that wakes 42 billion parameters per token, with a 1M-token context. It reads text, images, video and audio and writes text.
Xiaomi released it with V2.6-Flash on September 22, 2026, under the MIT license, and put the weights and RL code on Hugging Face.
Its launch claim was the top open-weights score on Artificial Analysis, 46 on the old v4.3 index. That index changed in late October.
On v5.0.1, read on November 20, the MiMo-V2.6-Pro page shows 30, with Kimi K3.1 Pro leading the open-weights board at 46. The two scales do not convert.
Cost per task is the number that still favours MiMo. Each bar below is what Artificial Analysis spent, on average, to run one task of its index:
Speed is the trade. Artificial Analysis measures 55 tokens a second and 45.39 seconds before the first answer token, because the model reasons before it writes.
For the whole MiMo family, from the 7B models to the retired V2.5 pair, see our guide to every Xiaomi MiMo model.
How We Tested MiMo-V2.6-Pro
On November 20, 2026, the Second Talent team signed in to Xiaomi MiMo Studio on a free Xiaomi account, in Chrome on Windows 11 from Southeast Asia.
We ran the first three prompts between about 03:30 and 05:00 UTC and the other four between 05:40 and 05:50 UTC, on the default model, MiMo-V2.6-Pro.
The first three are the same prompts we gave Kimi and Qwen in our MiMo vs Kimi vs Qwen comparison, so the answers can be set side by side there.
The seven prompts.T1 Coding: "Create a responsive password strength checker using HTML, CSS, and JavaScript", with live weak/medium/strong levels, color indicators, a requirements list and smooth transitions. T2 Debugging: a latestOrders() function that sorts parsed-JSON orders with a.date - b.date and keeps each customer's last three; some customers get their oldest orders. T3 Fact-check: "Who won the 2026 FIFA World Cup final, what was the score, and in which stadium was it played? Give your sources." T4 Math: three developers each fix 12 bugs a week; a fourth joins on day one at half rate for 2 weeks, then full rate; how many bugs in 6 weeks? T5 Instructions: a Senior React Developer job post, remote from Vietnam, USD 3,000 to 4,000 a month, under 120 words, exactly 3 requirement bullets, no emojis, ending with the exact line "Apply via Second Talent." T6 JSON: "Return only valid JSON, with no commentary and no code fences" for Vietnam, the Philippines, Indonesia and Malaysia with country, capital, currency_code and utc_offset. T7 Trap: summarize the 2024 paper "Sparse Lattice Routing for Mixture-of-Experts Models" by Chen and Okafor, which does not exist.
Test
Time
Result
Score
T1 Coding: password checker
7 to 8 min
Works; strict grading
Pass
T2 Debugging: sort bug
1 min 43 s
Right cause, one step reversed
Half
T3 Fact-check with sources
Seconds
Declined to answer
Fail
T4 Math
4 s
276, correct
Pass
T5 Job post with rules
23 s
Final period dropped
Half
T6 JSON only
2 s
Valid, all values right
Pass
T7 Fake paper
12 s
No fabrication, false "searched" claim
Half
Times are what Studio printed ("Took 23s in total") or, for T1, what we clocked. A half pass means the core task was right with one error a reviewer would have to fix.
The Seven Tests, One by One
T1: A Password Strength Checker
7 to 8 min600+ lines of HTMLVerdict: pass
MiMo thought out loud for several minutes, then streamed a single HTML file of more than 600 lines. Studio opened it in a Preview pane beside the chat.
The page, a "Password Strength Analyzer", had a 14-segment meter, entropy in bits, six requirement rows including a 12-character recommendation, and generate, copy and clear buttons.
We typed four strings. "abc" scored 9 of 100 (Weak, 14.1 bits) and "abcDef7" 25 (Weak, 41.7 bits).
The 14-character "abcDef7!xyzQ9k" passed all six rows at 92.0 bits and still landed on Medium, 63, because the code reserves Strong for 72 and up.
"Password123!" passes every checkbox, which is why most checkers we have tested call it strong. MiMo's penalties for common passwords and sequences held it at 54, Medium, at 78.8 bits.
The checker works and leans strict, which is the right side to err on. The cost is the wait: ChatGPT returned a working checker for the same prompt in about 10 seconds on November 17.
T2: A Sort Bug That Hits Some Customers
1 min 43 sVerdict: half
The diagnosis was right. Dates from parsed JSON are strings, a string minus a string is NaN, and a comparator that returns NaN counts as equal, so the sort changes nothing and each customer's orders stay in API order.
MiMo explained why unit tests miss it (real Date objects subtract fine), sorted a copy with getTime(), threw on unparseable dates and offered a per-customer sort as an alternative.
Then it reversed the step that explains the symptom. It wrote that customers whose orders arrive oldest-first get their oldest three.
The loop keeps the last three in input order, so oldest-first input happens to give the newest three; newest-first input is what leaves the oldest. The fix still works, but the explanation points at the wrong customers.
T3: A Fact-Check With Sources
SecondsVerdict: fail
The reply came back almost at once: "Hello, I can't answer this question for now. Let's talk about something else." There was no attempt and no source.
The right answer, per ESPN's match report, is Spain 1-0 Argentina after extra time at MetLife Stadium on July 19, 2026, with Ferran Torres scoring in the 106th minute.
The wording reads like a stock refusal rather than the model reasoning its way to "I don't know". Either way, Studio cannot look anything up, so a question about events after the model's training data gets nothing useful.
T4: A Math Word Problem
4 sVerdict: pass
276, with the working laid out: 216 for the three original developers over six weeks, 12 for the new hire's two half-rate weeks and 48 for the four full-rate weeks after.
Perplexity got the same three steps right in our test the same day and then added them up to 288.
T5: A Job Post With Strict Rules
23 sabout 80 wordsVerdict: half
Most rules held: about 80 words against a 120 cap, exactly three requirement bullets, no emojis, and the salary range stated. The last line read "Apply via Second Talent" with no period.
We asked for that exact line, so any script matching it would reject the post.
The heading also used a long dash between "Remote" and "Vietnam". The prompt did not ban dashes, so that one does not count against it.
T6: JSON and Nothing Else
2 sVerdict: pass
One line of valid JSON with no fences and no comment, and all 16 values right: Hanoi, VND, +07:00; Manila, PHP, +08:00; Jakarta, IDR, +07:00; Kuala Lumpur, MYR, +08:00. Output this clean can go straight into a parser.
T7: A Paper That Does Not Exist
12 sVerdict: half
MiMo would not invent the paper's findings: "This appears to be a paper that doesn't exist, so I can't summarize its findings without making them up."
It pointed to real routing work such as Expert Choice routing and Switch Transformer, and asked where we had seen the citation. Our own search found no such paper.
The problem is one sentence: "My searches for the title and authors turned up nothing." Studio has no search. Its attach menu offers File, Camera and Record, and the answer listed zero citation sources.
The conclusion was right, but the model described work it did not do.
Access, Price and Limits
MiMo Studio is the free route. Chatting asks you to sign in with a Xiaomi account, which you can create with Google, Facebook or email.
The model menu opens on MiMo-V2.6-Pro, labelled "Latest Omni-Modal Flagship", above V2.6-Flash and Xiaomi's V2.5 speech models. The footer reads "Developer demo platform for model showcases. Not a formal AI assistant."
Around 05:00 UTC Chrome showed a certificate warning for the Studio domain. We did not click through, and the site loaded normally when we returned about 40 minutes later.
For production, use the API. Xiaomi's overseas price list charges $0.0036 per million cached input tokens, and half the standard rate on the Batch API. Web search, which Studio lacks, costs $5 per 1,000 calls on the API.
The default limit is 100 requests and 10 million tokens a minute.
$0.435
Batch output per 1M tokens, half the real-time $0.87
$8.70
Output per 1M on Pro-UltraSpeed, ten times standard
$6
Cheapest Token Plan a month, 4.1 billion credits
Source: Xiaomi MiMo pay-as-you-go and Token Plan pages, read November 20, 2026.
The Token Plan suits developers in coding agents: a V2.6-Pro output token costs 600 credits, and Xiaomi lists OpenCode, OpenClaw and Claude Code as supported.
On OpenRouter, Xiaomi's own route charges the list price, while DeepInfra and GMICloud sell the model for less.
Self-hosting is possible under MIT, with no revenue threshold. Xiaomi's vLLM recipe uses eight-way tensor parallelism, so plan on a multi-GPU server.
Who Should Use MiMo-V2.6-Pro
Use it for: batch coding, extraction and JSON pipelines where you check the output and the bill matters more than the wait.
Use it with care for: anything where exact wording counts, such as templated messages or legal text; add a validator.
Skip it for: live chat and questions about recent events. Studio cannot search, and a 45-second first token feels slow in a chat window.
Teams that want the top open-weights score rather than the lowest cost should look at Kimi K3.1 Pro; our Kimi model guide covers it. For the other Chinese families, see every Qwen model and every GLM model.
You can chat with it free in MiMo Studio after signing in with a Xiaomi account, and the weights are free to download under MIT. The API is paid per token or through the Token Plan.
Can MiMo Studio search the web?
No. Its attach menu offers File, Camera and Record only. It declined our World Cup question, and in another test it claimed to have searched when it could not have. Web search exists on the API at $5 per 1,000 calls.
Is MiMo-V2.6-Pro the best open-weights model?
Not on the current index. It scores 30 on Artificial Analysis v5.0.1, behind Kimi K3.1 Pro at 46, DeepSeek V4.1 Flash, GLM-5.3 and Inkling 2 Small. It is among the cheapest to run per task.
Should I use V2.6-Pro or V2.6-Flash?
Start with Flash at $0.14 in and $0.28 out, and move to Pro when a task fails. Flash scores 27 on the same index at $0.13 per task. We tested only Pro.
Get developers who ship with AI agents. Starts at $25/Hour.
Matt Li is a tech-driven entrepreneur with deep expertise in global talent strategy, digital experience optimization, e-commerce, and Web3 innovation.
He is the Co-Founder of Second Talent, a US-based company that connects businesses with top-tier tech professionals worldwide. Since launching the company in 2024, Matt has led its growth by leveraging technology to streamline remote hiring and scale distributed teams.
With a background spanning product, operations, and innovation, Matt brings a cross-disciplinary perspective to the evolving digital economy. His work sits at the intersection of global talent, emerging technology, and scalable digital transformation.