Update: MiniMax M3.1 Flash lands on Together AI — eval flags a catch
A day after the weights dropped, Together AI hosts it and independent evals land: cheap and capable — but a ~93% hallucination rate, because it never abstains.
Copy markdown
Now hosted on Together AI — no GPUs required
The 428B/23B-active MoE MiniMax open-weighted on Oct 23 is now served in production by Together AI, so you can call M3.1 Flash through a managed endpoint instead of standing up 252 GB of weights yourself. The hosted API runs $0.15 per million input tokens, $0.60 output, and $0.015 cached.
The independent numbers are in
Artificial Analysis now scores M3.1 Flash at 42 on its Intelligence Index — up from M3's 29 — at roughly $0.22 a task. On DeepSWE coding it reaches 73.7% for under $0.30 a task, matching DeepSeek V4.1 Flash. For high-volume agent loops, that's frontier-flash behaviour at a floor price.
The catch: it almost never says "I don't know"
The same eval clocks a ~93% hallucination rate — not because it's weak, but because it attempts nearly every question instead of abstaining; factual accuracy on AA-Omniscience is 56%. Put retrieval and a verification pass in front of it before you trust it on anything factual.