MiniMax Drops Code 3.1 With a 1M-Token Flash Model and Five Reasoning Levels
MiniMax refreshed its coding agent with a new interface, the M3.1-Flash-Preview model, a five-level reasoning slider, and an upcoming M Plan subscription.
- MiniMax Code v3.1.0 ships with a redesigned interface and refreshed brand identity
- M3.1-Flash-Preview debuts as the default model, tuned for fast everyday coding
- New five-level reasoning slider (low, medium, high, xhigh, max) trades latency for depth
- 1M-token context window, but the model is gated to MiniMax Code, no public API
- Computer Use, MiniApp, Git Graph, custom themes and live agent action tracking upgraded
- Token Plan closes to new signups; M Plan takes over with same pricing and access to H3
MiniMax Code 3.1 adds a faster model and a five-level reasoning dial
MiniMax released version 3.1.0 of MiniMax Code on . The update combines a new coding model with expanded computer controls, clearer agent activity, repository tooling, interface changes, and a revised subscription plan.
Flash stays inside MiniMax Code
M3.1-Flash-Preview is a text model optimized for routine development work. MiniMax lists feature implementation, bug fixes, edge-case handling, regression tests, and change-impact analysis among its intended uses. The Preview is now the default model in the MiniMax Code picker.
| Attribute | Details |
|---|---|
| Context window | Up to 1 million tokens |
| Reasoning levels | low, medium, high, xhigh, and max |
| Primary role | Everyday coding tasks that benefit from fast responses |
| Availability | MiniMax Code only |
| Public API | No generally available endpoint announced |
A million-token context window can accommodate large repository snapshots, specifications, logs, and prior conversation history in one session. MiniMax has not published retrieval tests showing how reliably the model uses information across that full window.
The earlier M3 flagship, released in 2026, targets broader and more demanding reasoning workloads. Flash Preview has a narrower coding brief, and MiniMax has supplied no evidence that it matches M3 on architecture, benchmark performance, or multimodal capability.
Five settings tune speed against depth
The reasoning control adjusts how much processing M3.1-Flash-Preview applies to each turn without requiring a model change. Lower settings suit small edits and direct questions, while higher settings allocate more effort to debugging, multi-file changes, and complex implementation plans.
MiniMax has not disclosed latency, compute consumption, or quota usage for each level. Teams will need to measure those differences on their own workloads, especially at xhigh and max.
The agent exposes more of its work
Version 3.1.0 updates several parts of the coding environment:
- A refreshed visual identity and redesigned interface
- Improved Computer Use for browser and desktop automation
- MiniApp enhancements and custom themes
- A Git Graph panel for viewing repository history
- Real-time tracking of agent actions
Computer Use allows MiniMax Code to click, type, and navigate browser or desktop interfaces while completing a task. MiniMax ties this capability to the product’s broader M3 multimodal foundation. Its announcement describes Flash Preview as a text model and does not establish that Flash itself receives images or screen frames. The release notes also omit details about approval gates, permissions, and sandboxing for automated actions.
M Plan takes over subscriptions
MiniMax is replacing Token Plan with M Plan while keeping pricing and text-usage quotas steady. Token Plan is closed to new purchases. Existing subscribers with automatic renewal can remain on their current plan and migrate later.
Selected M Plan tiers will include access to the full MiniMax model lineup, including the H3 video model. The announcement does not identify which tiers receive that broader access.
From through , measured in UTC+8, daily logins earn double points. MiniMax says those points can be applied to code models, including M3.1-Flash-Preview, and to H3 video generation.
Anonymous tests point to MiniMax, with limits
MiniMax launched Flash Preview without a model card, benchmark report, standalone price, or public API terms. Days before the release, developers linked a free anonymous OpenRouter model named Space Bunny Alpha to an early build of the same system.
One independent test sent 24 token-count probes through OpenCode’s Space Bunny route, and all 24 matched MiniMax models tested on the same gateway. A separate measurement set reported 50 matches from 50 probes. These tests compare token-accounting patterns produced by a model’s tokenizer, which can help identify a model family. They cannot determine the exact checkpoint. MiniMax has not confirmed that Space Bunny Alpha and M3.1-Flash-Preview are the same model.
Test the preview on repository work
MiniMax Code users can evaluate the model immediately through the default picker. API-dependent applications and third-party agent frameworks currently lack a supported deployment path. A controlled evaluation can run identical tasks across all five reasoning levels and measure:
- Time to a correct, test-passing patch
- Response and task-completion latency
- Accuracy across large repository contexts
- Regression rates and unnecessary file changes
- Tool actions recorded in the live activity trace
- Quota consumption, where the interface exposes it
Published architecture details and SWE-bench results apply to the earlier M3 flagship. Flash Preview needs a dedicated model card and independent measurements before those figures can characterize its performance.
Comments
