MiniMax-M3.1-Flash is the flagship open-source language model in the M-3 lineup, built on a mixture-of-experts architecture with native multimodality. The model demonstrates strong capabilities in programming, agentic tasks, and cybersecurity. Its total number of parameters is about 428B, but due to sparse activation, only about 23B parameters are engaged per token, providing high performance with economical use of computational resources. The model supports a context window of up to 1,000,000 tokens, allowing it to work with large code repositories, long documents, and multi-step agentic sessions.
A key architectural innovation inherited from the M3 lineup and strengthened in version 3.1 is MiniMax Sparse Attention (MSA). MSA is a block-sparse attention mechanism built on top of Grouped Query Attention (GQA). It consists of two branches: the lightweight Index Branch evaluates all key-value blocks and independently selects the most relevant blocks (Top-k) for each GQA group, while the main Main Branch performs exact attention only over the selected blocks, always retaining the local (nearest) block for stability. This approach reduces attention computation and thereby significantly accelerates the prefill and decoding stages without sacrificing quality.
Compared with the previous version, MiniMax-M3, the 3.1-Flash model received three important changes that increase efficiency and performance. First, MSA sparse attention is now used in all layers of the model — in M3, the first three layers did not use sparse attention, which further reduces computational load. Second, quantization-aware training — the model officially ships in mxfp8 format, with experts in nvfp4, which reduces VRAM pressure while preserving quality. Third, support for DSpark speculative decoding. These architectural decisions make M3.1-Flash not merely a faster version of M3, but a qualitatively more efficient model for industrial deployment.
Among the key benchmarks, the developers compare their model not only with open but also with proprietary solutions. In BankerToolBench (agentic use of tools and APIs) — 76.2, the best result among those listed, above GLM-5.3 (74.5) and proprietary GPT-5.6 Sol (74.6). In DeepSWE v1.1 (solving real tasks in code repositories) — 73.7%, above Kimi K3 (67.5%) and GLM-5.3 (66.9%). In LOCA-Bench 256K (long-context work) — 60.3, also first place among those listed in the comparison. At the same time, in CyberGym (cybersecurity) 82.2 is a strong but not first result: proprietary GPT-5.6 Sol scores slightly higher — 83.6. The same is observed in Automation-Bench v1.0.6 (automation of multi-step tasks) — 62.8, in OSWorld 2.0 (agentic environment control) — 61.7, and in HLE without tools (complex intellectual tasks) — 43.4. Overall, these results allow M3.1-Flash to be positioned as a model of frontier-level quality.
The use cases for MiniMax-M3.1-Flash follow logically from its architectural features and testing results. In software development, the model can act as an autonomous AI assistant: analyze repositories, perform long-running refactoring, debugging, and documentation, and independently deploy and test code. In agentic workflows, it is capable of planning and executing multi-step tasks, managing tools and APIs, autonomously conducting research, and making decisions without human intervention. Thanks to its 1-million-token context, the model is ideal for analyzing long documents — scientific papers with formulas and graphs, legal contracts, financial reports, and complete system log files. In multimodal applications, it can help analyze video, extract information from complex visual materials, and create interactive systems that work with text, images, and video simultaneously. Finally, in cybersecurity, the model demonstrates strong results in CyberGym, making it suitable for threat analysis, vulnerability discovery, and automation of security tasks. Flexible adjustment of reasoning depth (the effort parameter from low to max) allows the model to be adapted to tasks with different latency and quality requirements.
| Model Name | Context | Type | GPU | Status | Link |
|---|
There are no public endpoints for this model yet.
An endpoint is a server with pre-installed software that automatically downloads neural network weights and launches an OpenAI-compatible API to handle incoming requests.
A private endpoint is created within the user's project, belongs exclusively to the user, and is billed per server configuration in accordance with the infrastructure pricing.
We recommend creating private endpoints when you need to:
| Name | GPU | TPS | Max Concurrency | |||
|---|---|---|---|---|---|---|
1,048,576.0 pipeline |
6 | $14.10 | 1.010 | Launch | ||
1,048,576.0 tensor |
8 | $18.35 | 1.258 | Launch | ||
1,048,576.0 tensor |
4 | $19.23 | 1.968 | Launch | ||
1,048,576.0 tensor |
4 | $19.23 | 1.968 | Launch | ||
1,048,576.0 tensor |
8 | $31.39 | 1.256 | Launch | ||
Contact our dedicated neural networks support team at support@immers.cloud or send your request to the sales department at sale@immers.cloud.