The Chinese laboratory has announced scores of 73.7 per cent on DeepSWE and 82 per cent on CyberGym; the model, tested anonymously under the name ‘Space Bunny’, will publish its weights at a later date.
MiniMax has launched M3.1 Flash, the new fast version of its model line, touted as a generational leap for agent tasks, programming, computer control, and cybersecurity. The Chinese startup reveals that a preview version was already circulating incognito under the name "Space Bunny" and processed more than 200,000 billion tokens. According to it, inference is nearly three times faster and the cost of using agent tasks is 70% lower than the previous model, with throughputs exceeding 200 tokens per second.
Claimed results at the level of the best
MiniMax is 73.7% ahead of DeepSWE v1.1 (standalone software development), 76.2% ahead of BankerToolBench, 61.7% ahead of OSWorld V2 and 29.6% ahead of Agents' Last Exam — the latter two placing it first among Chinese models — as well as 82% on CyberGym, a vulnerability discovery and validation test. The price is set at $0.15 per million input tokens, 0.60 output tokens and 0.015 for cached entries.
M3.1 Flash is accessible via the API, the M Plan subscription and the MiniMax Code application. The model's weights will be published at a later date, and a Pro version, geared towards maximum performance, is announced for later.
Read the full article on the original source
Go to the source →