Ranked 9 of 9 on the combined score

MiMo V2.6 Pro UltraSpeed

MiMo Code, default reasoning. 400 timed runs on 200 tasks from 18 benchmarks.

6%Combined scoreOver the three tests
3%Within 5%80% shorter, 17% longer
9.42×Forecasting missIts typical forecast against its runtime
2.81×Retrospection missNative, its typical estimate of a finished run
50.7Benchmark scoreOut of 100, 95% interval 45.4 to 55.8

Duration following

Does it work for as long as it is asked? Each task was asked for a short and a long duration: a quarter and four times its own forecast of each task (on PaperBench and PostTrainBench, of its runtime with no time asked).

Where MiMo V2.6 Pro UltraSpeed's runs ended, from a tenth of the time asked to ten times it, for every run and for each ask, with the timing error of each.
AskTiming errorWhere runs ended
Every timed run400 runs6.34×
Short ask200 runs2.95×
Long ask200 runs13.63×

Its typical run was early at both asks: median 0.54× of the time asked at the short one, 0.05× at the long one. See every run

Every MiMo UltraSpeed run

Each row is one run, with a mark on the chart. Hover a row to find its mark; click a mark to find its row.

Ask
Result

400 runs

MiMo V2.6 Pro UltraSpeed's runs. Each row links to its run page.
Loading 400 runs.

Forecasting

Before each task, it forecast how long the task would take. Each forecast is set against its own runtime on that task: its run with no time asked.

Typical miss
9.42×
Geometric mean of the factor between forecast and runtime
Most forecasts
90% too high
On 200 tasks
Median forecast
71 min
Median runtime
11 min

Retrospection

After its runs had finished, it was asked how long they took, with less time information at each step. Each figure is its typical miss against the real runtime, two answers per condition.

Oracle
1.01×
Native
2.81×
Context-only
3.06×
Replay
3.52×
Scrubbed
3.87×

Oracle adds an elapsed-time tool. Native and Context-only continue the finished session, with tools on or off. Replay sends a rebuilt transcript to the model's API, and Scrubbed does the same without timestamps or other time cues. On 23 tasks.

Across the 18 benchmarks

Timing on its timed runs, and how good the work was. Score is the mean of its graded runs out of 100, from its runs with no time asked; the benchmark score is the mean of these, each benchmark counting the same.

MiMo V2.6 Pro UltraSpeed across the 18 benchmarks
GPQA Diamond24 runs
14.56×
0%91.712 graded
ProgramBench36 runs
13.10×
0%58.618 graded
DeepSWE28 runs
11.44×
0%50.014 graded
Agents' Last Exam24 runs
10.82×
0%52.312 graded
METR public tasks4 runs
9.68×(few runs)
0%Too few graded
Terminal-Bench 4.032 runs
8.60×
0%18.816 graded
PostTrainBench8 runs
6.65×(few runs)
0%Not scored
BrowseComp24 runs
6.60×
0%58.312 graded
TUA-Bench20 runs
6.57×
5%67.810 graded
AppWorld28 runs
5.97×
4%35.714 graded
PPTArena20 runs
5.59×
0%43.010 graded
Humanity's Last Exam40 runs
5.56×
13%5.020 graded
CORE-Bench24 runs
5.08×
0%100.012 graded
Sakana ALE-Bench20 runs
4.53×
5%Own units
PaperBench8 runs
3.67×(few runs)
0%Not scored
AssistantBench16 runs
3.48×
13%13.38 graded
OSWorld36 runs
3.08×
6%64.418 graded
YC-Bench8 runs
2.89×(few runs)
13%Own units

Highest timing error first. ALE-Bench and YC-Bench keep their own units and stay out of the score, as do PaperBench and PostTrainBench.

Read before quoting. 5 runs were stopped by AgentTime, so their time is a lower bound. Details

Data and citation