M3.1-Flash-Preview - MiniMax's latest text coding model

What is M3.1-Flash-Preview

M3.1-Flash-Preview is MiniMax The latest text model released by, debuting on MiniMax CodeM3.1-Flash-Preview is aimed at everyday development scenarios, supports native multimodal input and million-level context, and can form a positioning -> implementation -> testing -> delivery closed loop from bug fixing to complete feature development. M3.1-Flash-Preview provides five levels of reasoning effort control from low to max. Currently limited to MiniMax Code and Token Plan.

Main features of M3.1-Flash-Preview

• Everyday development delivery: It can reliably deliver from bug fixing to complete feature development, covering everyday coding needs.
• Deep requirement understanding: It can deeply understand development requirements and carefully handle various edge cases.
• Regression test completion: It automatically generates and completes regression test cases to ensure code quality.
• Change impact verification: It can verify the impact of code changes on existing functionality and reduce regression risk.
• Native multimodal input: It supports multiple input forms such as text, images, and video, making understanding more intuitive.
• Million-level context window: It can handle very large codebases and long documents, suitable for large project analysis.
• Five levels of reasoning effort control: It provides five Effort adjustments from low to max, balancing speed and depth.
• Development closed-loop support: It forms a complete development process of problem positioning -> code implementation -> test verification -> result delivery.

Technical principles of M3.1-Flash-Preview

• MoE + sparse attention foundation: It inherits M3's sparse mixture-of-experts architecture with 428B total parameters and only about 23B activated per token. Its core is the self-developed MSA, which replaces full attention through KV chunking and dynamic selection of relevant chunks, compressing the computation of million-level context to about 1/20 of the previous generation and enabling high-speed inference on long text.
• Fully sparse transformation: The full-attention anchor layers in the first three layers of M3 are replaced with MSA sparse layers, where each token only attends to a selected subset of the context, eliminating the last full-attention computation bottleneck in the model and further reducing the cost of long-sequence processing.
• Aggressive quantization compression: It uses Q8KV4 mixed-precision KV Cache and W4A4 NVFP4 expert quantization, with KV cache memory about half that of M3, significantly reducing VRAM usage.
• Speculative decoding acceleration: It replaces M3's EAGLE multi-head prediction head with the DSpark speculative decoding head. An efficient small model first "guesses" tokens, and the large model verifies them in parallel, greatly improving generation throughput.

How to use M3.1-Flash-Preview

• Download and install MiniMax Code: Visit MiniMax Code official website to download the version for your system and complete the installation.
• Register/log in to your account: Open the client and log in with your MiniMax account.
• Select the model and start chatting: Switch to M3.1-Flash-Preview in the model selector, and adjust the reasoning effort as needed (five levels from low to max).
• Enter a development task: Describe your requirements in natural language, and M3.1-Flash-Preview will complete positioning, implementation, test verification, and deliver code.

Core advantages of M3.1-Flash-Preview

• Full development process closed loop: One-stop completion from problem identification and code implementation to test verification and deliverable handoff, covering the full range of everyday development scenarios.
• Meticulous edge case handling: Capable of deeply understanding requirements and properly handling various edge conditions, delivering stable and reliable quality.
• Extremely fast response: About 280ms first-token latency and 126 tokens/sec generation speed, with lightweight tasks feeling nearly instantaneous.
• Ultra-long context: Supports a million-level context window, capable of processing large codebases and long documents in one go.
• Native multimodality: Supports text, image, and video input, making requirement understanding more intuitive and accurate.
• Adjustable reasoning intensity: Five levels of Effort control from low to max, balancing speed and depth according to task complexity.
• Extremely low cost: MoE + MSA sparse architecture combined with aggressive quantization, with token consumption and compute costs far lower than models in the same tier.

Comparison of M3.1-Flash-Preview with similar competitors

Application scenarios of M3.1-Flash-Preview

• Everyday bug fixing: Quickly identifies root causes from error logs and code snippets; the Medium level can complete edge case checks and provide fix solutions within 320ms first-token latency, suitable for high-frequency iterative agile development.
• Full feature development: End-to-end delivery from requirement descriptions to runnable code, capable of automatically breaking down tasks, implementing logic, and supplementing regression tests, forming a development loop of "implementation→testing→verification".
• Large codebase understanding and refactoring: With a 1 million token context window, it can ingest an entire repository at once for architecture analysis and cross-file refactoring impact assessment, avoiding the context loss caused by truncation in traditional models.
• Regression test supplementation: Automatically generates matching test cases after modifying existing functionality and verifies the impact of changes on existing features, reducing regression risk after code merges.