MiniMax released its next-generation lightweight model M3.1 Flash on the evening of October 19, confirming that Space Bunny, an anonymous model that had been quietly tested on aggregate platforms, was an early version of the new release.
The company said M3.1 Flash delivers major gains in complex agent tasks, computer operation, software development and cybersecurity, with online inference speed nearly three times faster and usage costs cut by 70 percent. Output speed reaches more than 200 tokens per second in daily scenarios, with a median above 170 tokens per second thanks to a more aggressive sparse attention architecture.
On benchmarks, the model scored 73.7 percent on DeepSWE v1.1 for long-chain software engineering, 76.2 on BankerToolBench for multi-step financial workflows, 61.7 on OSWorld V2 and 29.6 on Agents' Last Exam, both first among Chinese models, and 82 percent on CyberGym for vulnerability discovery and verification.
Pricing was cut sharply: 1 yuan per million input tokens, 4 yuan per million output tokens and 0.1 yuan per million cached tokens, in line with DeepSeek's off-peak rates. The release came just two days after DeepSeek updated its V4.1 Flash model on October 17, giving China's flash-model segment a 72-hour one-two punch and signaling that lightweight models are moving from budget alternatives to mainstays for complex agentic work. A Pro version pursuing higher intelligence limits is coming, the company said.
Source: Shanghai Securities News / National Business Daily / China Foundation News, October 19, 2026.
