MiniMax-M2

MiniMax-M2 is an open-source large language model released by AGI startup MiniMax on October 27, 2025 — the follow-up to M1 (June 2025) and, per 绿洲资本, purpose-built for agent and coding scenarios, pitched as a rare balance of intelligence, cost and speed: roughly 8% of Claude Sonnet's price at nearly 2x inference speed . Within MiniMax's model matrix — LLMs (the M series), video (Hailuo) and speech/music models — M2 anchored the company's full-modality push into its Hong Kong IPO, which closed subscription in early January 2026 at over 1,209x oversubscribed . The line has since moved fast: an M2.1 upgrade landed in December 2025 (reaching fourth on LMArena's WebDev board, behind only Anthropic, OpenAI and Google ), and by June 2026 M2 was superseded by M3, which added a 1M context window, native multimodality and top-tier SWE-Bench scores . Note that elsewhere's corpus on M2 itself is largely investor-amplified (绿洲资本, 云启资本, both MiniMax backers), so the framing skews promotional.