Loading page…
The model is available through MiniMax’s API and coding tools. Its speed and benchmark gains remain company-reported, while downloadable weights are promised later.
Listen to this story
The October 19, 2026, release positions M3.1 Flash around the economics of multi-step agent work: MiniMax claims task-level usage costs 70% less than with its predecessor, alongside nearly threefold faster online inference. The cost reduction applies to agent tasks, not necessarily to the model’s per-token prices, and the capability evidence consists of company-reported benchmarks rather than independent validation. Developers can use the model through MiniMax API, M Plan and MiniMax Code; downloadable weights are promised later, with no release date set.
Published rates are $0.15 per million input tokens, $0.60 per million output tokens and $0.015 per million cached input tokens.
MiniMax reports scores of 73.7% on DeepSWE v1.1, 61.7 on OSWorld V2 and 82% on CyberGym.
MiniMax says everyday output can exceed 200 tokens per second, with median speeds consistently above 170.
MiniMax released M3.1 Flash on October 19, 2026, pairing lower-priced AI agent work with claimed gains in coding, computer operation and cybersecurity. The company says task usage costs are 70% below its predecessor’s and online inference is nearly three times faster. Access is live through MiniMax’s services; downloadable model weights will follow later.
The launch offers two different cost measures. MiniMax publishes flat rates for input, output and cached input, all charged per million tokens. Separately, it says the overall usage cost for agentic tasks—work involving multiple steps and tool calls—is 70% lower than its predecessor’s. That is a claim about running tasks, not a stated 70% reduction in each listed token rate.
MiniMax’s listed rate for input supplied to M3.1 Flash.
MiniMax’s listed rate for output generated by the model.
MiniMax lists cached input as a separate, lower-priced category.
On speed, MiniMax says everyday output can exceed 200 tokens per second, with median speeds consistently above 170. These figures describe output generation, while the company’s capability tests examine whether the model can carry out work. MiniMax attributes its inference-efficiency gains to more aggressive sparsity, and describes the overall online inference improvement as nearly threefold.
MiniMax’s benchmark results span several kinds of work rather than one general score. The company says M3.1 Flash ranks among the top performers on AutomationBench-AA, which tests complex task automation. It also describes enterprise workflows across software services, including sales, marketing, operations, customer service, finance and HR. Those are MiniMax’s capability claims, not independently validated outcomes.
Cybersecurity adds a different measure: MiniMax reports 82% on CyberGym, which it connects to vulnerability discovery and validation. Its internal security experts also evaluated code auditing, vulnerability patching, penetration testing, malware analysis and reverse engineering. The published CyberGym score and those internal assessments are distinct pieces of the company’s case for using the model in security work.
M3.1 Flash is available through the MiniMax API, M Plan and MiniMax Code. The announcement directs developers to its API platform and offers a desktop download for MiniMax Code. These are the launch’s immediate access routes. MiniMax says model weights—the files needed to run the model separately—will be released later, without giving a release date.
The release also identifies an earlier anonymous version. MiniMax says Space Bunny was an early M3.1 Flash model and processed more than 200 trillion tokens while operating anonymously. That disclosure connects the named launch to an earlier deployment, but the volume figure describes usage; it is separate from the benchmark results offered to support the model’s capabilities.
Story updates
Together AI’s model card describes MiniMax M3.1 Flash as a 428-billion-parameter model that activates just 23 billion parameters per token. It also details changes to attention, decoding and developer controls behind MiniMax’s October 19, 2026 rollout. The useful distinction is between the model’s underlying design and the settings developers can adjust when calling it.
M3.1 Flash retains MiniMax M3’s core mixture-of-experts architecture. That design distinguishes the model’s total parameter count from the smaller portion activated for each token. Together lists roughly 428 billion parameters in all, with 23 billion active. The larger figure therefore describes the full model, not the amount activated at every step.
The changes sit inside that inherited structure. Together identifies revisions to both attention—the model component used to work with its context—and decoding, which produces its output. These are specific differences from M3, rather than a claim that MiniMax replaced the entire architecture. The model card also describes training changes aimed at reducing the memory needed to serve it.
Together says quantization-aware training covers the routed experts and attention paths. This trains those components with reduced numerical precision in mind. The stated benefits are a smaller memory footprint and more efficient serving. Those are infrastructure claims about running the model, separate from whether it successfully finishes a coding or computer-use task.
For applications, the model’s input and output capabilities are not symmetrical. It accepts text, images and video, but returns text. Together lists a context window of up to one million tokens—the material the model can work with in a request. Its suggested uses include long agent sessions, large codebases and lengthy documents.
Together lists access on both serverless and dedicated infrastructure, using the model identifier MiniMaxAI/MiniMax-M3.1-Flash. Requests authenticate with a Together API key. This is an access route for the released model; the listing does not date the start of Together’s integration.
Reasoning depth: set reasoning_effort to low, medium, high, xhigh or max.
Sampling: Together recommends temperature 1.0 and top-p 0.95.
MiniMax’s launch announcement supplies the performance case. The company reports 73.7% on DeepSWE v1.1, which tests software work including understanding code, making changes and validating results. It reports 61.7 points on OSWorld V2, a test of computer operation and tasks across applications. These are MiniMax’s reported results, not independent verification of the architectural changes.
MiniMax also claims nearly threefold faster online inference and 70% lower overall usage costs for agent tasks than its predecessor. That task-cost claim is distinct from the published token rates: $0.15 per million input tokens, $0.60 for output and $0.015 for cached input. Together lists the same rates.
Hosted access and downloadable model files also remain distinct. In its October 19 announcement, MiniMax said M3.1 Flash was available through its API, M Plan and MiniMax Code, while model weights would be released later. That was a dated release plan, not confirmation that downloadable weights accompanied the rollout.
minimax.ioLoading discussion...
Join the conversation
Explain which measure would help you judge value more clearly.
Be the first to share a perspective or an experience.
Reader comments
Newest comments first. Replies stay oldest first.