MiniMax releases open weights for M3.1 Flash, a 428-billion-parameter model, to compete with Llama and Qwen

What was released

MiniMax, a Chinese AI startup, published the open weights of M3.1 Flash, its latest language model, on October 23, 2026. The weight release came four days after the model went live as a hosted service, on October 19. Now any developer can download the files and run the model on their own infrastructure, without relying on the company's hosted API.

M3.1 Flash is a mixture-of-experts (MoE) model with roughly 428 billion total parameters, but it activates only 23 billion per processed token. That's the same efficiency logic used by other heavyweight open models, such as DeepSeek-V3 and Alibaba's Qwen3 family: the model is large enough to hold a lot of knowledge, but computationally cheaper to run than a dense model of the same size.

Technical specs

A context window of up to 1 million tokens puts M3.1 Flash in the long-context model tier, useful for anyone who needs to feed the model entire codebases, lengthy documentation, or long agent sessions without truncating the history.

Built for coding, agents, and computer use

According to MiniMax, M3.1 Flash was designed with three scenarios in mind: code generation and review, building AI agents, and computer-use tasks, a category of models trained to operate graphical interfaces on their own (clicking, typing, navigating) instead of just responding with text. It's the same ground where Anthropic, OpenAI, and Google have been investing with their own computer-use models, but here with open weights available for download.

For anyone building agents, the combination of long context with multimodal input (text, image, and video) matters in practice: an automation agent can receive a screenshot, a log snippet, and the entire conversation history in the same call, without needing separate pipelines for each input type.

How to run the model

MiniMax published deployment guidance for two inference frameworks widely used by teams already running open models in production: SGLang and vLLM. Both are optimized to serve large LLMs with low latency and already support MoE architectures, which makes it easier to adopt M3.1 Flash without rewriting the serving layer from scratch.

The model also has support via FlagOS for multiple chip platforms. In practice, this means dependence on Nvidia GPUs to run M3.1 Flash is lower than in many Western releases, a detail that isn't trivial for Chinese companies (and for any team looking to diversify inference hardware) given export restrictions on advanced chips.

Efficiency techniques behind the model

Three optimizations support the promise of running a 428-billion-parameter model without the inference cost of an equivalent dense model:

• Sparse attention: reduces the computational cost of attention over long sequences, which is exactly what makes the 1-million-token window viable without blowing up response time.
• Training-aware quantization (QAT): the model is already trained anticipating that it will run at reduced precision, which preserves more quality than quantizing a finished model after training.
• Speculative decoding: a smaller model generates token candidates that the main model validates in batch, speeding up text generation without changing the architecture.

These are techniques well known from the inference optimization literature, but combining all of them in a single open-weights release is what sets M3.1 Flash apart from MiniMax's own previous versions.

Where this fits in the open-model race

The market for open weights used in production already has established competitors: Meta's Llama family and Alibaba's Qwen family dominate much of the integration work in developer tools, coding agents, and self-hosted assistants. M3.1 Flash enters this race with a specific argument: a low activation ratio (23B out of 428B) combined with a 1-million-token context and native support for video as input, something not every open competitor offers today.

Based on the material released so far, the exact license under which the weights were made available and direct comparative benchmarks against Llama or Qwen on coding and agent tasks remain open questions. Without that data, it's up to whoever evaluates the model to run their own tests before swapping out their production stack.

For teams that already use SGLang or vLLM as their serving layer, the evaluation path is straightforward: download the weights, spin up a local endpoint, and compare latency, cost per token, and output quality against the model currently in use, before deciding on a migration.