MiniMax Launches M3.1 Flash MoE Model With 1M Context and New Reasoning Defaults

MiniMax has released M3.1 Flash on its application programming interface, introducing a mixture-of-experts model designed for agentic tasks. The Chinese artificial intelligence firm announced the launch alongside pricing details and a timeline for open weights. This release follows earlier reporting by Threat Times on October 19 regarding the model's enhanced capabilities.

The model contains 428 billion total parameters, with approximately 23 billion active parameters per inference. It supports a context window of one million tokens and accepts text, image, and video inputs. According to MarkTechPost, the API charges $0.15 per million input tokens and $0.60 per million output tokens. MiniMax confirmed that Space Bunny was an earlier build of this same model.

View the original post on X

Technical Specifications and Performance

M3.1 Flash targets coding, office automation, and visual reasoning. MiniMax reports that the model achieves near-frontier scores on agentic benchmarks. On Automation-Bench v1.0.6, the model scored 62.2, leading Opus 5 and GPT-5.6 Sol. It also topped LOCA-Bench 256K with a score of 60.3. These performance figures are self-reported by MiniMax, with no independent replication confirmed yet.

Coding results are mixed compared to competitors. The model scored 73.7 on DeepSWE v1.1, slightly above GPT-5.6 Sol but below DeepSeek-V4.1-Flash-1017. Terminal Bench 4.0 yielded the lowest result at 38.1, significantly trailing Opus 5’s 51.8. SWE Refactor Bench scores were also lower than Opus 5. Vision benchmarks showed strong performance on BabyVision with tools at 91.0.

Integration Changes for Developers

A major change for users is that reasoning is always on and cannot be disabled. Developers must choose from five effort levels: Low, Medium, High, Xhigh, and Max. MiniMax recommends Max as the default setting. Requests that attempt to disable reasoning or set effort to none will return an error. This breaks compatibility with some existing integrations.

Integrations that parse inline thinking tags from previous versions need updating. Reasoning now arrives in a separate reasoning_content field, which may be absent. MiniMax has not published a technical report detailing specific architectural changes in the Flash variant. Model weights are due for release on October 23, 2026, enabling self-hosting.

What should you do?

Developers using MiniMax APIs should review their integration code for hardcoded reasoning disable flags. Applications relying on inline tag parsing must be updated to handle the new separate reasoning field. Users should test the new effort levels to balance cost and performance. Monitor for the October 23 weight release if self-hosting is required.

What to do and how to stay safe: MiniMax

• Audit API calls for parameters that disable reasoning or set effort to none, as these will now error.
• Update parsers to handle the new reasoning_content field instead of looking for inline thinking tags.
• Test the model with different effort levels to determine the optimal balance for your specific workload.
• Prepare infrastructure for self-hosting if you plan to use the model weights upon their October 23 release.

Step-by-step guide: Prompt Injection Warning Signs: How to Spot AI Manipulation

General security guidance from the Threat Times newsroom. It is not confirmed advice from the organisations named in this story.

Frequently asked questions

Is the reasoning process optional in MiniMax M3.1 Flash?

No, reasoning is always on and cannot be disabled. Developers must select an effort level instead.

When will the model weights be available for self-hosting?

MiniMax has scheduled the release of model weights for October 23, 2026.

How does M3.1 Flash perform on coding benchmarks compared to Opus 5?

It performs worse on Terminal Bench 4.0 and SWE Refactor Bench, but slightly better on DeepSWE v1.1.