88Z AI, 23 October — MiniMax has open-sourced the weights for its MiniMax M3.1 Flash model. Featuring 428 billion total parameters and 23 billion activated parameters, it supports a context window of one million tokens. Within 72 hours of launch, it processed over 100 trillion tokens, with inference costs cut by more than half compared to the previous generation. It ranks first among domestic models in several Agent benchmarks. Ten chip manufacturers, including Huawei Ascend and NVIDIA, along with inference framewor
At a glance
- MiniMax M3.1 Flash is open-sourced with halved inference costs and 1M token context support
- Ten chip vendors and major inference frameworks completed adaptation on day one to boost ecosystem adoption
The update
MiniMax has officially open-sourced the weights for its MiniMax M3.1 Flash model to accelerate the adoption of AI technology. The model features 428 billion total parameters and 23 billion activated parameters, supporting a context window of up to one million tokens for complex long-text tasks. Since its release, it has demonstrated strong performance, processing over 100 trillion tokens within just 72 hours. Meanwhile, inference costs for the new generation have dropped by more than 50% compared to the previous version, significantly improving cost-efficiency. In multiple Agent benchmarks, the model ranks first among domestic offerings, showcasing superior logical reasoning and task execution capabilities. To build a robust open-source ecosystem, ten major chip manufacturers, including Huawei Ascend and NVIDIA, along with prominent inference frameworks such as vLLM and SGLang, completed technical adaptations on the launch day, ensuring developers can deploy and use the model quickly.
Sources & editorial note
This article is an AI-assisted synthesis by 88Z AI based on the public sources below. Product claims and availability are attributed to their publishers. It does not represent independent testing or first-hand reporting.
- MiniMax 稀宇科技Source published
Source dates above refer to the original announcements; the date under the headline is this article’s publication on 88Z AI.