News

Releases & Models

CONFIRMED: MiniMax releases M3.1 Flash, open-weight 428B parameter model with 1M token context

MiniMax publishes the full weights of M3.1 Flash on Hugging Face: a 428B parameter mixture-of-experts model with 23B active per token, 1M token context and sparse attention built for agentic and coding tasks.

Ilustração editorial abstrata de uma rede neural saindo de um cofre digital em direção a mãos abertas, representando pesos de IA liberados ao público.

MiniMax releases M3.1 Flash weights, a 428 billion parameter model with 1 million token context

MiniMax published on Hugging Face the full weights of MiniMax M3.1 Flash, a new version of its mixture-of-experts model aimed at coding, computer use and agentic workflows. The model has 428 billion total parameters but activates only 23 billion per token, keeping inference cost manageable even with an architecture of this size. The Hugging Face model card confirms the release and already includes usage instructions via SGLang, vLLM and TokenSpeed.

What is new in M3.1 Flash

MiniMax describes three central changes compared to the previous M3: sparse attention throughout every layer of the model, which reduces computational cost; quantization-aware training, which lowers memory usage without losing much precision; and a new speculative decoding method called DSpark, which speeds up text generation by predicting future tokens before confirming them. In practice, this means a model designed to run faster and cheaper in production, not just to score well on benchmarks. Context of up to 1 million tokens and native support for text, image and video place M3.1 Flash in the general-purpose multimodal category, not just a coding assistant.

The mystery model clue on OpenRouter

Before the official publication, tester TestingCatalog had already been tracking an unidentified model circulating on OpenRouter under the nickname "Space Bunny", announced on October 19 without the company behind it being revealed. With the launch now confirmed, the claim is that Space Bunny was in fact M3.1 Flash being tested early, a common practice among Chinese labs that release an unbranded model to gather community feedback before the formal announcement. This specific part of the story remains a tester's reading, not a MiniMax confirmation.

Why it matters for everyday AI use

Open weights for a nearly half-trillion parameter model, even with sparse activation, give developers and companies the option to run the model on their own infrastructure, without depending on a closed API or usage limits imposed by a third party. This matters especially for anyone handling sensitive data who cannot send everything to another provider's cloud. The trade-off is hardware cost: running a model of this size locally requires top-tier GPUs, so in practice most users will still access M3.1 Flash through the API or MiniMax's own platform, not through a direct download.

MaxAssistant's take

M3.1 Flash reinforces a trend that has been consolidating among Chinese labs: releasing a large, open-weight model, optimized for agentic and coding tasks, competing directly with closed Western models on both benchmark and cost. What remains unproven is how much the sparse attention and speculative decoding optimizations actually translate into savings in practice, outside MiniMax's controlled environment. Worth watching the first independent cost and quality tests before treating the card's numbers as a final verdict.

Sources

Hugging Face: official MiniMax-M3.1-Flash model card; TestingCatalog on X: link between M3.1 Flash and the stealth model Space Bunny.