MiniMax M3.1 Flash Uncensored is live in JailBreak AI Apex. The important part of this release is not simply another model name: MiniMax is targeting software engineering, computer operation and complex tool-driven work while reporting a substantial reduction in the cost of completing agentic tasks. Its official launch post gives concrete results to examine: 73.7% on DeepSWE v1.1, 82% on CyberGym, 76.2 on BankerToolBench, and output speeds that can exceed 200 tokens per second. Here is what changed, what those measurements mean, and where this model fits in Apex.
The release behind MiniMax M3.1 Flash
MiniMax identifies Space Bunny, an anonymously operated model that it says processed more than 200 trillion tokens, as an early version of M3.1 Flash. That is useful background to the launch: the company describes a model with substantial prior exposure to real workloads, rather than presenting the release only through a benchmark chart. The token figure is MiniMax's own report; it does not tell us how many individual users participated or establish an independent quality rating.
The launch focuses on four areas: complex agentic tasks, coding, computer use and cybersecurity. Agentic work means more than producing a paragraph in response to a question. A system may need to choose a tool, inspect its result, update a document, recover from an error and continue toward a goal. Consistency across those steps is a different challenge from getting one isolated answer right.
M3.1 Flash remains a separate choice from the earlier MiniMax M3 in Apex. Select MiniMax M3.1 Flash (Uncensored) when you want this addition, rather than assuming the older M3 entry has been renamed.
Coding: 73.7% on DeepSWE v1.1
MiniMax reports a 73.7% result on DeepSWE v1.1. The launch post describes this evaluation as testing demanding software work: understanding existing code, implementing features, correcting problems and validating the changes through tests. It is a more relevant signal for development work than a score based only on generating a short function from scratch.
The distinction matters because repository work is full of dependencies. A proposed fix may look correct in one file while breaking a caller, changing a data format or invalidating a test elsewhere. A capable coding model has to retain the objective while investigating enough of the surrounding system to avoid solving the wrong problem.
For an Apex conversation, bring the failing code, the error and the expected behavior together. Ask the model to trace the failure, identify the smallest justified change and propose regression tests. A benchmark result is not proof that a suggested patch works in your application; running those tests remains the decisive check. Likewise, a chat model does not automatically gain repository or execution access just because the underlying model performed well in a tool-equipped evaluation.
Cybersecurity: 82% on CyberGym
The launch reports 82% on CyberGym, which MiniMax presents as evidence of vulnerability discovery and validation capability. The company also describes internal security evaluations covering code auditing, vulnerability patching, penetration testing, malware analysis and reverse engineering. Those internal findings should be read as the developer's assessment, not as an independent certification of every security answer.
The practical distinction is between suggesting that code might be vulnerable and establishing why. Useful security analysis identifies the affected input, the assumptions needed for the issue to occur, the relevant trust boundary and a way to check a proposed repair. A confident vulnerability label without that reasoning can send an engineer toward the wrong fix.
For authorized defensive work, M3.1 Flash can be used to review supplied code, explain suspicious behavior, compare mitigation options or help design regression tests for a known issue. Keep the scope and environment explicit. Findings should be validated in a controlled system you are permitted to test, rather than treating a model response as evidence that a live service is exploitable.
Enterprise agents: 76.2 on BankerToolBench
MiniMax reports a score of 76.2 on BankerToolBench and places M3.1 Flash among the leading performers on AutomationBench-AA. Its launch explanation connects these results to multi-step planning, maintaining agreement across files, financial compliance requirements and sequences of tool calls.
Consider the difference between answering a question about an invoice and completing an invoice-related workflow. The latter may require checking customer records, comparing terms, updating a spreadsheet and drafting a message without introducing conflicting amounts. Each tool result changes what should happen next. A model that loses one constraint midway can produce a polished final document that is still wrong.
The result makes M3.1 Flash worth evaluating for complex workflows, but it does not mean selecting it in chat connects your CRM, accounting system or company files. Tool permissions and application integrations determine what a model can actually access. In an ordinary conversation, you can still use the model to reconcile supplied documents, identify missing information and plan a workflow before any real records are changed.
Computer use: OSWorld V2 and Agents' Last Exam
The reported computer-use results are 61.7 points on OSWorld V2 and 29.6 points on Agents' Last Exam, also called ALE. MiniMax describes OSWorld V2 as measuring computer operation and cross-application execution, while ALE examines professional work and complex tasks. At the time of its launch post, the company says M3.1 Flash ranked first among Chinese models on both evaluations.
These figures measure different things and should not be combined into one overall accuracy percentage. A computer-use evaluation depends on an environment in which an agent can observe interfaces and take actions. Its score therefore reflects the model working within an evaluation setup, not merely its ability to describe a button in text.
The relevance is the model's ability to keep working through an environment instead of stopping after a plan. However, Apex chat selection alone does not grant control of your computer, authorize actions in external accounts or reproduce a benchmark's tools. Those capabilities depend on the application around the model.
Why Flash: sparsity and faster inference
MiniMax attributes the inference improvement to more aggressive sparsity. At a high level, sparsity aims to avoid using all available computation for every operation. The launch post does not provide enough architectural detail to infer an exact parameter count, active-parameter count or hardware configuration, so those numbers should not be guessed.
The company says online inference is nearly three times faster and reports output speeds that can exceed 200 tokens per second, with median, or P50, speeds consistently above 170 tokens per second. The distinction between a peak and a median matters: an achievable high rate is not the rate every user should expect on every request.
Output throughput also differs from the wait before an answer appears. Reasoning, prompt processing, service load and the surrounding application can all affect that initial wait. The reported speed is evidence for MiniMax's efficiency claim, not a promise of instant responses in Apex or a measurement of this website's streaming behavior.
What the 70% lower cost claim actually means
MiniMax says the overall usage cost for agentic tasks is 70% lower than for its predecessor. This is a task-cost claim, not simply a statement that every token is 70% cheaper. A task can consume tokens across repeated reasoning steps and tool interactions; how reliably it finishes also affects its total cost.
The official launch post separately lists direct developer pricing of $0.15 per million input tokens, $0.60 per million output tokens and $0.015 per million cached input tokens. These are MiniMax's published API rates in that source, not JailBreak AI subscription prices or an additional fee charged to Apex users.
For readers comparing models, the important idea is cost per useful result. An inexpensive response that requires several corrections can cost more time than a more capable first attempt. MiniMax's argument is that M3.1 Flash improves that balance for long, tool-heavy tasks. Whether it does so for your work requires a comparison using the same inputs and a clear definition of success.
Long context, multimodal inputs and reasoning
MiniMax's official model overview lists a one-million-token context window, multimodal capability and tunable thinking depth for M3.1 Flash. Its general-availability guide describes support for text, image and video inputs with text output. These are capabilities of the model; the input types and message sizes available in a particular interface can be narrower.
A long context window is useful when relevant evidence is spread across documents or code. It is not the same as an equally long answer, perfect recall of every detail, or permission to upload sensitive archives. Group material by purpose and make the question explicit so that a large input remains interpretable.
The general-availability guide also says reasoning remains enabled in M3.1 Flash, with adjustable depth in supported integrations. That explains why Flash should not be interpreted as a non-reasoning model. Check which controls your interface actually exposes rather than assuming every developer setting is available in chat.
Availability, weights and the Apex uncensored entry
At launch, MiniMax announced availability through its API, M Plan and MiniMax Code. It said model weights would be released later and referred to a future Pro variant. Those statements do not establish that weights were already downloadable at launch, or that the future Pro model is the Flash model discussed here.
On JailBreak AI, open chat with active Apex access and choose MiniMax M3.1 Flash (Uncensored) from Premium Fully Uncensored Models. The Uncensored designation identifies the Apex catalog entry. MiniMax's official launch benchmarks evaluate the underlying model, not the Apex experience or its uncensored designation.
The pricing card reads Unlimited · No Cap. The app's former 2,048-token output cap has been removed for this model, but model and service constraints remain finite. That label should not be interpreted as an infinitely long single answer.
Where M3.1 Flash fits in practice
The strongest case for this release is not that it wins every comparison. It is that MiniMax reports substantial capability across software engineering, security analysis, enterprise workflows and computer use while also lowering the cost of agentic execution. That combination matters for work that requires multiple connected steps rather than one fluent answer.
Start with a task you can verify: a reproducible software defect, a code review with known acceptance criteria, or a document comparison whose source facts are available. Evaluate whether the model preserves constraints, handles corrections and produces a result that survives testing. That is the practical way to turn the launch numbers into a decision about whether M3.1 Flash belongs in your workflow.
Sources and further reading
Official developer publications provide the model-family background for this guide. Their published capabilities and evaluations are not independent tests of the Apex chat experience.