MiniMax M3.1 Flash

	MiniMax M3.1 Flash(minimax.io)
	510 points by jtg 6 days ago \| 218 comments
	tl;dr: MiniMax released M3.1 Flash, a next-gen model (previously tested anonymously as "Space Bunny") that claims competitive performance with top models on agentic, coding, computer use, and cybersecurity benchmarks like DeepSWE (73.7%), OSWorld V2 (61.7), and CyberGym (82%). It runs at 200+ TPS with pricing at $0.15/$0.60 per 1M input/output tokens—70% cheaper than its predecessor. Available now via API, with open weights to be released later.
	HN Discussion: ↓Model is benchmaxed and underperforms in real-world use despite impressive benchmark claims ↓LLM announcements are generic marketing with cherry-picked benchmarks and need independent verification ↑The price-to-performance ratio of new cheap models is genuinely impressive and disruptive •Technical analysis of pricing/weights details, confirming claims are plausible or noting open weights potential ~Mixed real-world testing results: strong on some benchmarks but slow/token-hungry in practice