SYLIN

MiniMax M Flash

by MiniMax · Model · Language & code models

promising

Ranks second among open-weights models on Arena documents; 3.1 is still unverified

M3.1 Flash, upgraded from the Space Bunny stealth model, is available by API; weights promised later

Why we say so

Arena, an independent measure, puts M3 Flash second for document understanding among open weights (1437), behind Kimi K2.6 (1450) and just ahead of Gemma 4 (1434). The gaps are small. For 3.1, only MiniMax's own claims exist: nearly matching DeepSeek on agent tasks and being cheapest on AutomationBench. Its weights are promised but not out.

Promised Oct 25: weights released later

Versions

VersionReleasedRank
MiniMax M3.1 FlashLATEST
unproven Only MiniMax's claims so far; API only, weights promised
Oct 25
MiniMax M3 Flash
promising Second among open-weights models on Arena documents
#2 documents (open weights)

Evidence

  1. Nov 17
    Pipeline-parallel prefill (PP2) sustains 23% more requests per GPU than DEP2 at a 2s TTFT target
  2. Nov 9
    #2 on Arena document among open weights (1437)
  3. Oct 25
    Almost as good as the latest DeepSeek model, and the most cost-efficient on AutomationBench
  4. Sep 13
    #2 on Arena document among open weights (1435)

Cost to run

Alternatives

Kimi#1 on Arena document among open weights (1450)
Gemma#3 on Arena document among open weights (1434)
Inkling-SmallTrails GLM 5.2 and Kimi K3 on Artificial Analysis and is far slower than MiniMax M3 and GLM 5.2; strength is audio and vision