MiMo V2.6 Pro open weights tested
Eighth on the extended CymraegBench leaderboard with an overall score of 80.73.
The facts
| Developer | Xiaomi |
|---|---|
| API ID | mimo-v2.6-pro |
| Test service | Local OpenAI-compatible endpoint |
| Reported context | 262,144 tokens |
| Test cost | $0.00 (local service) |
CymraegBench result
| Overall score | Welsh reasoning | Practical Welsh | Translation BLEU |
|---|---|---|---|
| 80.73 | 83.79 | 77.67 | 61.98 |
The same sample was used as for the other leading models: up to 100 cases from each of the 19 core suites, with random seed 1. The 232 CCC cases do not contribute to the overall score.
MiMo V2.6 Pro is eighth on the extended leaderboard: 0.36 points behind Qwen3.8 2.4T-A95B and 0.92 points ahead of Qwen3.8 Flash. It completed 2,036 of the 2,053 cases; the remaining 17 invalid outputs scored zero.
Note on long outputs: 55 requests failed to produce a final answer within the initial 2,048-token limit. Only those requests were rerun with higher limits. Thirty-eight were recovered, including eight at 16,384 tokens. Seventeen remained invalid: 11 correction-and-explanation cases, four place names, and one case each in ARC Challenge and ARC Easy. There were no unresolved infrastructure errors.
CCC Language Audit
| Subset | Score |
|---|---|
| Correction and explanation | 70.70 |
| Idiom | 81.25 |
| Terminology | 78.13 |
| Summarisation and quotation | 87.89 |
| References | 74.69 |
| Welsh prompt | 100.00 |
| English prompt | 100.00 |
| Average across the seven subsets | 84.66 |
On correction and explanation, MiMo scored 75.00 for correction and 42.19 for explanation. Eleven of the 64 cases were invalid outputs counted as zero. It completed the rest of the CCC audit.
See the full leaderboard, CCC Language Audit, or download the summary, scores by suite, CCC dimensions and the metadata.