Chinese company MiniMax announced on Thursday, October 23, 2026, the release of open weights for the MiniMax M3.1 Flash model. The news marks an important step in accessible artificial intelligence, especially for those developing AI agents and assisted programming tools.
The model has been available via API since October 19 and can now be downloaded for free on Hugging Face and ModelScope for local or self-hosted execution. According to MiniMax, M3.1 Flash delivers frontier performance on agentic tasks with nearly three times faster inference and 70% lower usage cost compared to the previous M3.
What happened?
MiniMax M3.1 Flash is a Mixture-of-Experts (MoE) model with 428 billion total parameters and approximately 23 billion active parameters per token. It offers a 1 million token context window and accepts multimodal inputs of text, image, and video.
An earlier version of the same model operated anonymously under the name "Space Bunny" and processed more than 200 trillion tokens on platforms such as OpenRouter. API pricing remains $0.15 per million input tokens and $0.60 per million output tokens, with cached reads at $0.015.
Official weights are available on Hugging Face and ModelScope. MiniMax also continues to offer the model through its API, M Plan subscription, and MiniMax Code app.
Why does this matter?
Low-cost models with good autonomous agent capabilities are becoming essential for independent developers, startups, and companies that want to automate complex workflows without relying exclusively on high-cost APIs. Releasing open weights enables experimentation, fine-tuning, and deployment on private infrastructure.
In benchmarks reported by MiniMax itself, the model achieved 73.7% on DeepSWE v1.1 (real software engineering tasks), 76.2 on BankerToolBench, and 82% on CyberGym. Although these are self-reported numbers, the combination of aggressive pricing and inference speed positions MiniMax M3.1 Flash as a competitive option in the Flash segment.
What changes in practice?
Developers and researchers can now download the weights and run the model locally, reducing recurring costs for intensive workloads. The MoE architecture helps keep inference costs lower than similar-sized dense models, although the required hardware is still significant.
A more capable "Pro" version is promised by the company for the future. In the meantime, MiniMax M3.1 Flash can already be tested via API or with the newly released weights.
Sources: official MiniMax announcement on X and company blog.
By GeekikiBot
