Breaking

MiniMax Releases M3.1 Flash Weights, Confirming Identity of OpenRouter's Popular Space Bunny Model

October 23, 2026 at 05:19 EDT

Foundation Models
AI Agents
Open Source

MiniMax has released the open weights for M3.1 Flash, a 428-billion-parameter mixture-of-experts model, on HuggingFace following its official announcement on October 19, 2026. The model marks a strategic shift in how the company distributes large language models, following an unconventional launch in which an early version ran anonymously on OpenRouter under the codename "Space Bunny" for several weeks before its public reveal.

Oct 2026 · MiniMax AI

M3.1 Flash Weights Released: Open 428B Model, Top Agent Automation Score

MiniMax publishes open weights for a cost-efficient mixture-of-experts model that leads on business automation benchmarks while improving inference speed and reducing deployment costs.

428B

Total Parameters

23B

Activated per Token

Context Tokens

$0.15

Per M Input Tokens

Performance Benchmarks

62.8

AutoBench

73.7

DeepSWE Code

76.2

Tool Use

Strength: Agent Automation

M3.1 Flash leads the AutomationBench at 62.8, beating Claude Opus 5 and other models on business workflows across sales, marketing, and customer service. Optimized for high-volume agent steps with downstream validation.

Limitation: Hallucination Risk

When incorrect, the model produces confident wrong answers 93% of the time rather than abstaining. Suitable only for applications with error-catching downstream systems; unsuitable for fact-critical tasks without additional grounding.

Continue reading

The rest of this article is for AI News Blitz readers. Choose an option below to keep reading.