The model's weights will be released on Oct 23, and the vLLM project has already shipped Day-0 support for the architecture. During its stealth phase, MiniMax used performance feedback to cut error rates to 0.7% and increase throughput to 100 tokens per second. While the model matches or halves the peak rates of DeepSeek V4.1 Flash and beats it on Automation-Bench, it trails the DeepSeek model in four of six comparable benchmarks.
Wednesday, October 21, 2026
Story
MiniMax Sets Oct 23 Weight Release for M3.1 Flash AI Model
AI2 days ago
The M3.1 Flash model from MiniMax previously operated under the name 'Space Bunny,' a stealth project that spent 10 consecutive days as the top model on OpenRouter and processed 56.7 trillion tokens. This official version runs at 3x the speed of the M3 model at one third the cost, with flat rate API pricing of $0.15 per 1 million input tokens and $0.60 per 1 million output tokens. Performance benchmarks include a 74% score on DeepSWE for coding and 91% on BabyVision for visual reasoning.
Looking for more context or analysis?
Tilt chat provides reliable, tailored insights, alerts, and automations.