MiniMax, the Chinese AI developer, released M3.1 Flash on 19 October, an update to the M3 model it launched in June. The company says agent-style tasks now cost 70% less than on its previous model and that responses arrive close to three times faster. It also confirmed that an anonymous model called Space Bunny, which had been showing up on OpenRouter since late September, was an early build of M3.1 Flash. By MiniMax's count Space Bunny processed more than 200 trillion tokens before the reveal.
The pricing is aggressive. The API costs $0.15 per million input tokens and $0.60 per million output tokens, with cached input at 1.5 cents. Output speed is above 200 tokens a second at everyday peak, with a median over 170. The model accepts text, image and video, handles prompts up to a million tokens, and offers five levels of reasoning effort from Low to Max. Its developer guide lists 428 billion parameters, about the same as M3, so the speed is credited to more aggressive sparsity rather than a smaller model.
MiniMax published benchmark scores that include 61.7 on OSWorld V2, 76.2 on BankerToolBench, 82% on CyberGym and 73.7% on DeepSWE v1.1, and says no Chinese model has beaten its OSWorld V2 result. These numbers come from the vendor and have not been reproduced independently. Tech Insider also points out that the 70% saving is measured against MiniMax's own earlier generation, not against rivals' prices.
The model is available through the MiniMax API, its M Plan subscription and the MiniMax Code app, where a preview appeared on 27 September. Requests to the older M2.5 and M2.7 models are being redirected to M3.1 Flash, and those two retire on 18 November.
MiniMax describes the model as part of its ongoing exploration of how to balance efficiency against capability. The release lands in a crowded week for cheap Chinese models. MiniMax listed in Hong Kong in January after a $619 million IPO, and its shares fell after DeepSeek launched V4.1-Flash in September. DeepSeek says its updated model solved 77% of DeepSWE v1.1 tasks against 73.7% for M3.1 Flash, and MiniMax's input price matches DeepSeek's off-peak rate.
About MiniMax
MiniMax is a Shanghai AI startup that went public in Hong Kong in January 2026. Caixin reported that it filed its prospectus on 31 December 2025, offering about 25.4 million new shares at HK$151 to HK$165 each, to raise up to HK$4.2 billion (about $540 million). At the top of that range the company was valued at about HK$50.4 billion, and trading was due to start on 9 January.
1
Works credited
2
Models indexed
Models on AIPopDB
Two things are still to come. Open weights for the Flash model are promised for later, though Tech Insider says MiniMax gave no date and an unverified search summary points to 23 October. A more capable Pro model is also on the way, according to SiliconANGLE and ToolNavs.
Sources · 3