MiniMax launches M3.1 Flash, a 428B MoE model for agentic tasks

Chinese AI developer MiniMax made its new language model, M3.1 Flash, available on October 19, 2026. Built on a mixture-of-experts (MoE) design, the model has 428 billion parameters in total and is aimed at agentic workloads.

According to the details released, M3.1 Flash activates roughly 23 billion of those parameters for any given step. It also supports a context window of 1 million tokens, enough to take in long documents, large codebases or extended multi-step task histories in a single pass.

Pricing and open weights

MiniMax is charging $0.15 per million input tokens and $0.60 per million output tokens through its API. The company said the model's weights will be published on October 23, 2026. The licence under which they will be released was not specified in the available information.

An MoE model is split into many specialised sub-networks, and only a fraction of them run for each input. The aim is to keep the knowledge capacity of a very large model while bringing inference cost and latency closer to that of a much smaller one, which fits with the low price MiniMax has set.

Why it matters

Agentic systems call tools, write code and carry out multi-step tasks on a user's behalf, and they tend to burn through large volumes of tokens. That makes per-token pricing a deciding factor for developers building on top of such models.

Shanghai-based MiniMax is one of several Chinese companies that have drawn attention for both language and video generation models. Large models from Chinese labs with openly released weights have become a notable source of price pressure in the global market.

Benchmark results and comparisons with competing models were not included in the information available. Independent testing after the weights go public should give a clearer picture of how the model performs in practice.