MiniMax launches M3.1 Flash, a cheap agent model, with weights due 23 October
Sorted AI20 OctChinese AI company MiniMax has released M3.1 Flash, a low-cost model aimed at agent tasks, coding and computer use. MiniMax prices it at $0.15 per million input tokens, $0.60 per million output tokens and $0.015 per million cached input tokens, and says it will release the open weights on 23 October, though it has not yet published the licence. MiniMax reports strong scores on its own benchmarks, including 76.2 on BankerToolBench and 73.7% on a coding test, DeepSWE. Independent tests by Artificial Analysis, reported by VentureBeat, place it among the leading cheap models for intelligence but flag a reliability problem: it answered 56% of factual-knowledge questions correctly, yet when it did not know an answer it almost always produced a wrong one rather than admit uncertainty, which the benchmark counts as a 93% hallucination rate.
Why it matters: If you want a cheap model for high-volume agent or coding work, this is one to watch once the weights are out, but the hallucination finding means you should add checks or human review before trusting it on facts.
MiniMax ↗AI-assisted, reviewed by Sorted AI
▲ 0
0 comments