Xiaomi update lifts MiMo-V2-Flash tool-calling to 97%
Xiaomi's February refresh of MiMo-V2-Flash reports 78.6 on SWE-Bench Verified and much more reliable tool calls in thinking mode, along with shorter reasoning, though some benchmark scores went down.
Xiaomi has published release notes for a February 4 update to MiMo-V2-Flash, labelled mimo-v2-flash-0204. The page is titled as a launch of three new "MiMo-V2.6" models, but the text it contains covers only this point update. It gives no details about the V2.6 models.
The headline changes are in thinking mode. SWE-Bench Verified rose to 78.6, up from 74.2 in the January 12 build. Tool-call success climbed from 64% to 97.0%, and the original release scored 44%. Xiaomi says it has fixed the stability problems that hurt agent workloads. Arena-Hard (Hard Prompts) in thinking mode reached 60.6, and AA-IFBench rose from 64 to 72.
On factuality, the AA-Omniscience non-hallucination rate went from 9% to 52%. Accuracy on the same benchmark fell from 27 to 19. That pattern suggests the model now declines to answer more often, rather than knowing more.
Xiaomi also changed how the model generates its chain of thought, cutting average output length by 13% to 30%. Some benchmarks lost accuracy in exchange:
| Benchmark | Old acc. | 0204 acc. | Length cut | |---|---|---|---| | AIME25 | 94.8 | 91.1 | 30.0% | | HMMT Feb 25 | 94.2 | 92.9 | 26.7% | | LiveCodeBench-AA | 83.2 | 84.9 | 14.7% | | GPQA-Diamond | 83.7 | 83.8 | 13.9% |
Compared with the 0112 build, some scores fell. Non-thinking Arena-Hard Hard Prompts dropped from 52.7 to 49.3, and thinking-mode Creative Writing dropped from 90.4 to 85.8.
The API call method and model name are unchanged.