Chat model

MiMo-V2.6-Flash high · chat

high reasoning · 1 question per request

Accuracy, all 2,198 questions
72.1%70.2% to 73.9%, 95% interval; 1,585 right
Answers kept at 99% confidence
18 of 2,1981 wrong; share right 74.2% to 99.0%, 95% interval
Cost, $ per 1,000 questionsper 1M questions
$2.219$2,219Basis and price under Cost
Latency
21.4 smedian request, 1 question per request

Back to the leaderboard.

How we called it

Kind
Chat model
Access
OpenRouter chat API, Xiaomi's own endpoint (fp8), fallbacks off, output cap 65,536 tokens, reasoning included
Paid as
Provider service
Model ID
xiaomi/mimo-v2.6-flash
Parameters
Mixture of experts: 309B total, about 15B active per step. Size class uses the total (309B).mixture-of-experts backbone, as Xiaomi states it; source: huggingface.co/XiaomiMiMo/MiMo-V2.6-Flash-RL; read 9 November 2026
Confidence from
Written estimate
Run
877, 8 November 2026

Settings

Reasoning
high
Request format
1 question per request
Requests at once
32
Temperature
not recorded
Run record: harness seed, harness commit and prompt version
Harness seed
0
Harness commit
3e3f2d2446be
Prompt version
254597b8be16

Results

QuestionsAccuracy95% intervalRight
All questions72.1%70.2% to 73.9%1,585 of 2,198
Reading86.6%83.8% to 89.0%568 of 656
Grammar66.0%63.6% to 68.3%1,017 of 1,542

Comparison group: xyzAB. Compare with another setup.

Invalid answers: 64; read from broken replies: 1; empty replies asked again: 12 (6 came back with answers, 6 answers). Still invalid: 3 still empty when asked again, 61 cut off at the output limit. Without the rule: 71 invalid. An empty reply is asked once more; in a broken reply, any letter that is clearly readable is scored and the rest count as wrong. Invalid answers count as wrong (how).

With “Bilmiyorum” (run 879): Invalid answers: 53; read from broken replies: 1; empty replies asked again: 4 (4 came back with answers, 4 answers). Still invalid: 53 cut off at the output limit. Without the rule: 58 invalid. (how)

Accuracy by topic
TopicPartAccuracyRight
Word and phrase meaningsReading92.4%121 of 131
Sentence meaningsReading89.0%137 of 154
How a writer explains and develops ideasReading65.7%46 of 70
Topic and main point of a paragraphReading92.9%91 of 98
How a paragraph fits togetherReading79.2%84 of 106
Details that support a paragraph’s main pointReading91.8%89 of 97
Word types, such as nouns and verbsGrammar74.3%104 of 140
Groups of words built around a nounGrammar80.0%52 of 65
Verbs and their formsGrammar70.3%64 of 91
Verb forms used like nouns or modifiersGrammar76.4%55 of 72
Word endings and what they doGrammar83.5%66 of 79
How words are builtGrammar46.4%32 of 69
The roles of words and phrases in a sentenceGrammar67.6%98 of 145
How verbs show who acts and who receives the actionGrammar78.4%58 of 74
Types of sentencesGrammar71.8%61 of 85
Questions combining several grammar topicsGrammar54.1%119 of 220
Sound patterns and changesGrammar31.2%24 of 77
SpellingGrammar62.4%83 of 133
PunctuationGrammar62.4%73 of 117
Problems in sentence wordingGrammar73.1%128 of 175

English

The Text tab's 547 kept reading questions translated into English, asked with this setup's Text tab settings. Every setup: English reading; against Turkish: English and Turkish.

Right answers
487 of 547, first English run (878)
Accuracy
89.0%
Non-hallucination with “I don't know”
1.7%: 1 “I don't know” of 59 not right (run 880)
Answered with “I don't know”
99.8%: 546 of 547
Error rate with “I don't know”
10.6%: 58 of 546 answers
Cost, $ per 1,000 questionsper 1M questions, first English run
0.298298at list prices
Latency, median request, first English run
6.67 s; 8 requests at once
English run
878
Machine, English runs
hosted API
Date
8 November 2026
Request format
1 question per request
Reasoning
high
Prompt version
5c3a4849752e
Translations, SHA-256
bf0d47cd385a93050c65906977e199ef7e45d5848dbbdef2bc8d124f4d5abf68

Confidence

Answers kept at 99% or more
18, of which 1 wrong (share right: 74.2% to 99.0%, 95% interval); 2,180 skipped: 2,116 with the model less than 99% sure and 64 without a valid answer
Calibration gap
7.1 points
AUROC, all questions
0.71 0.68 to 0.73
AUROC, reading
0.78 0.72 to 0.83
AUROC, grammar
0.66 0.63 to 0.69
Brier score
0.177
Accuracy of its surest answers
most confident fifth: 91.8% right; most confident half: 85.9% right; most confident four fifths: 79.6% right
Typical score of a right answer
85.0%
Typical score of a wrong answer
75.0%

Repeats

We did not repeat this setup.

Latency

Latency, median request
21.4 s (this run)
Latency, 90th percentile
533.6 s
Requests at once
32
Throughput
7.7 questions per minute of request time
Request time
283.8 minutes, over the requests whose answers were kept

Latency: the time from sending one request to receiving its answer, measured from our computer, network and any retries within the request included. Request time: the time at least one request was in flight, over the requests whose answers were kept. It is not the run's elapsed time: it leaves out requests that failed and any pause before a run was resumed.

Cost

Cost per 1,000 questionsper 1M questions
$2.219$2,219
Whole run
$4.88
Basis
Estimated from published prices and recorded token usage
Price
$0.14 input, $0.0028 cached input, $0.28 output per 1M tokens (source, 2026-11-08)
Input tokens per question
370
Output tokens per question
7,785