MiniMax M3.1 Flash Preview Ships as Gated Reasoning Coding Model

MiniMax-M3.1-Flash-Preview is the latest M-series language model from MiniMax for agentic reasoning, tool use, coding, and long-context tasks, with tunable thinking depth and a 1,000,000-token context window.

MiniMax released the model on Sept. 27, 2026, in a controlled rollout that keeps access inside the company's Token Plan subscription and MiniMax Code tool, according to MiniMax's API documentation and BenchLM.ai's model page. The release is notable less for where the model sits in the M-series line than for how it is packaged: reasoning is always on, returns its reasoning content separately, and can be tuned across five effort levels. MiniMax's official account described the model as faster and lighter, built for teams running high-volume, latency-sensitive workloads.

How does MiniMax-M3.1-Flash-Preview fit into MiniMax's M-series lineup?

MiniMax-M3.1-Flash-Preview is the latest M-series language model, positioned for agentic reasoning, tool use, coding, and long-context tasks, per the company's API documentation. It follows the base MiniMax-M3 model, which MiniMax reported at 59.0% on SWE-Bench Pro, a benchmark for software engineering agent performance, according to Startup Fortune. The Flash Preview label separates the new model from that base release while keeping it inside the same product family.

MiniMax has kept the new model out of its broad API catalog for now. Instead, access runs through Token Plan and MiniMax Code, a pairing that ties the model to the company's own development tools. BenchLM.ai lists the release date as Sept. 27, 2026, and describes the announcement as being made for MiniMax Code.

What's new in MiniMax-M3.1-Flash-Preview?

The headline specification is a 1,000,000-token context window, according to MiniMax's API documentation. That lets developers pass large codebases, multi-file diffs, or long tool-call histories into a single prompt. For agentic coding workloads, the larger window reduces the need for external retrieval and keeps more of the execution context in the model's immediate attention span.

The model is natively multimodal, accepting text, image, and video inputs. That makes it capable of processing screenshots, UI captures, or video demonstrations as part of a coding or agentic task, rather than requiring a separate vision model in the pipeline. MiniMax's documentation lists text, image, and video as supported input types.

Reasoning behavior is the other major change. Thinking is enabled by default and cannot be disabled, according to MiniMax's documentation. MiniMax returns the reasoning content separately from the final response, which gives applications a clean way to surface or log the model's internal deliberation without mixing it into the answer.

How does MiniMax-M3.1-Flash-Preview handle reasoning content?

Separating reasoning content from the final answer is a meaningful design decision. Applications can store the reasoning content for auditability, feed it into downstream evaluation, or display it in a debugging view without polluting the user-facing response. MiniMax's documentation says the reasoning content is returned separately, which implies the API exposes a distinct reasoning field rather than embedding it in the message text.

For teams that run agentic loops, the separation also prevents the model's internal deliberation from being re-ingested as tool output. That reduces the risk of the agent confusing its own reasoning with external results, and it gives developers a clean boundary between the model's private deliberation and the actions it takes.

How does the effort parameter work on MiniMax-M3.1-Flash-Preview?

The model exposes an effort parameter with five settings: low, medium, high, xhigh, and max. The default is max, according to MiniMax's documentation. Developers can lower the setting to trade reasoning depth for speed, which is directly relevant to the high-volume, latency-sensitive workloads MiniMax cited in its announcement.

1. low applies minimal reasoning depth for high-throughput requests.
2. medium and high provide intermediate levels for typical agentic and coding workloads.
3. xhigh increases reasoning depth for complex multi-step tool use.
4. max, the default, applies the fullest thinking depth the model supports.
5. All settings keep thinking enabled; the model does not offer a no-reasoning mode.

The default effort level is max, the highest setting listed in the documentation. Developers who want faster responses must override the default. That distinction matters because the Flash Preview name implies speed, but the out-of-the-box behavior is deliberately maximized reasoning depth.

MiniMax-M3.1-Flash-Preview specifications at a glance

Attribute	Detail
Release date	Sept. 27, 2026
Availability	Token Plan and MiniMax Code
Context window	Up to 1,000,000 tokens
Thinking mode	Enabled by default; cannot be disabled
Effort settings	low, medium, high, xhigh, max
Input modalities	Text, image, video
Primary use cases	Agentic reasoning, tool use, coding, long-context tasks

Why does a 1,000,000-token context window matter for coding models?

Long-context support matters for agentic coding workloads because the model must retain repository structure, prior tool outputs, and conversation history in a single prompt. A 1,000,000-token window allows MiniMax-M3.1-Flash-Preview to ingest large codebases or long execution logs without external retrieval. The practical effect is that teams can give the model more of the codebase before asking it to plan a change.

For coding and agentic tasks, native video and image input lets the model observe UI behavior, error screens, or visual demonstrations rather than relying on text descriptions. The documentation lists text, image, and video as supported inputs, making the Flash Preview a multimodal tool rather than a text-only code model. That combination of long context and multimodal input is what MiniMax emphasizes in the documentation.

What is the significance of the Flash Preview naming?

The Flash Preview designation carries two signals. Flash models are typically lighter and faster variants of a base model, and preview status signals that MiniMax is still gathering feedback before finalizing the release. The official announcement reinforced the speed message by describing the model as faster and lighter for high-volume workloads.

The Flash Preview name follows a pattern in the M-series family that includes the base M3 model. MiniMax's public documentation does not position the new model as a successor to M3; instead, it lists M3.1 Flash Preview as the latest M-series language model. That distinction matters because the Flash label usually signals a lighter, faster variant, while the Preview label signals that the release is not final.

What does the gated launch mean for MiniMax's enterprise customers and the broader market?

The gated distribution is a strategic choice. MiniMax's official account announced the model's availability on the Token Plan in a post on X, describing it as faster, lighter, and built for teams running high-volume, latency-sensitive workloads. The announcement emphasized that existing Token Plan subscribers can access the model without extra setup.

For developers, the practical consequence is that MiniMax-M3.1-Flash-Preview is not yet available as a standalone API endpoint. Teams that want to evaluate the model must already be on Token Plan or working inside MiniMax Code, which narrows the early adopter pool to existing MiniMax customers. The gated approach also means public benchmark reports and failure analyses will lag the actual deployment of the model.

The model's positioning inside MiniMax's own tools suggests the company is using the release to harden the model against real usage before broader distribution. Preview releases in this format give the vendor control over the feedback loop while still generating production traffic from its most committed users.

MiniMax-M3.1 Flash Preview is now live on the Token Plan! Faster, lighter, and built for teams running high-volume, latency-sensitive workloads, now available under your existing Token Plan subscription, no extra setup required.

What has MiniMax said about the release?

MiniMax's public statement on X was short and operational. The company told Token Plan subscribers that the model was live, positioned it for high-volume workloads, and emphasized that no additional configuration was needed. The documentation adds technical detail: the model is the latest M-series language model, supports effort-based thinking tuning, and carries the 1,000,000-token context window.

BenchLM.ai's model page corroborates the Sept. 27, 2026 release date and lists the model under its MiniMax Code announcement. The page also tracks the model's public specification, including the context window and multimodal inputs. Together with the official X post and the API documentation, these sources give a consistent picture of a deliberately narrow initial release.

What does the SWE-Bench Pro score tell us about MiniMax's coding strategy?

The 59.0% SWE-Bench Pro score for the base M3 model, as reported by Startup Fortune, is a reference point for MiniMax's coding ambitions. SWE-Bench Pro evaluates how well a model can resolve software engineering tasks in realistic repository environments. The score belongs to the base M3 model, not to the Flash Preview, and MiniMax has not published a separate benchmark result for the new model in the documentation reviewed for this report.

For teams comparing models, the score provides one data point for M3's engineering-agent capability. The Flash Preview's value proposition is less about raw benchmark leadership and more about packaging reasoning depth, long context, and multimodal input inside MiniMax's own tools. Early adopters will have to measure the new model's quality for themselves.

How does the launch fit into China's broader AI model race?

MiniMax is one of several Chinese AI vendors shipping new models at a rapid pace, according to Startup Fortune's framing of the release. The company's decision to bundle the model with its own Token Plan and MiniMax Code product is part of that competitive pattern, where vendors use proprietary distribution to differentiate.

The gated release also gives MiniMax a distribution advantage inside its own ecosystem. Rather than competing for API traffic on price alone, the company can direct early demand toward its coding tool and subscription plan, where the model's long context and reasoning controls are showcased.

What should developers expect next from MiniMax's M-series?

MiniMax has not announced a general availability date for the Flash Preview outside Token Plan and MiniMax Code in the sources reviewed for this report. The model is designated a preview, which means its specifications and behavior can change before a wider release. The company's documentation is the only authoritative source for current behavior.

For now, the model is available to the developers who are already inside MiniMax's ecosystem. Those teams can test the 1,000,000-token context window, the five-level effort parameter, and the separately returned reasoning content. Teams outside that ecosystem have no direct access path until MiniMax widens the release.

MiniMax's base M3 model scored 59.0% on SWE-Bench Pro, a benchmark for software engineering agent performance, according to Startup Fortune.

How should teams evaluate MiniMax-M3.1-Flash-Preview?

Evaluation should start with the model's default behavior at max effort, then test lower effort levels to understand the latency-quality curve. Teams should also verify how the separately returned reasoning content interacts with their own logging and safety systems. Because the model is gated, evaluation will initially depend on access through Token Plan or MiniMax Code.

The key variables to measure are context retention at the 1,000,000-token boundary, coding accuracy on repository-scale tasks, and response latency across the five effort settings. Those three measurements will tell teams whether the Flash Preview delivers on MiniMax's positioning as a faster, lighter model for high-volume workloads.