MiniMax M3.1 Flash

Built to be the worker, not the boss

Most model launches are pitched as the smartest model yet. MiniMax M3.1 Flash is more useful to think about as something else: a capable, inexpensive model you can run thousands of times a day inside an agent system without the bill getting out of hand.

MiniMax released a preview on September 27, 2026, made it generally available on October 19 through the MiniMax API, its M Plan subscription, and its MiniMax Code coding tool, and published the weights on Hugging Face and ModelScope on October 23. It is designed for coding, agents, and computer-use tasks.

Specs

MiniMax says it used quantization-aware training and speculative decoding to make inference cheaper, and the weight release includes a small drafter model for the speculative decoding. The repository is around 252GB, with deployment guidance for SGLang and vLLM. Like other MoE models, only 23B parameters do work on each token, but the whole model has to be loaded, so self-hosting still needs multi-GPU hardware.

What independent tests show

Artificial Analysis, which runs its own evaluations, gives M3.1 Flash a score of 43 on its Intelligence Index, ranking it near the top of its price class. It measured output speed at about 96 tokens per second, slightly below the median for comparable models, and time to first token around 1.3 seconds.

Two details in that independent data matter more than the headline score. First, the model produced roughly 180 million output tokens across the index, against a median of about 100 million. It thinks out loud a lot, and those reasoning tokens are billed as output. Second, Artificial Analysis lists it as the cheapest model per task in its comparison, at about $0.22 per index task, so the low token price more than covers the extra verbosity on those tests. At least one comparison drawing on independent leaderboards put DeepSeek’s newest Flash model ahead on quality, so cheapest per task does not mean best.

MiniMax publishes its own agent and coding benchmark results as well. Treat those as vendor-run until independent results agree.

Where it fits

M3.1 Flash makes sense as the worker tier in a model routing setup:

• Subagents that read, search, and summarize for a planner model, as in planner-worker agents
• Repetitive coding tasks such as writing tests, fixing lint errors, or applying the same change across many files
• Extraction and classification over large document sets, where the 1M context and flat pricing help
• Research subagents that gather sources before a stronger model writes the conclusion
• High-volume agent workflows where cost per run decides whether the workflow is viable

It is a weaker choice for difficult architecture decisions, high-stakes reasoning, or anything where one wrong answer is expensive. Those are the cases for model escalation to a frontier model, with an evaluation gate deciding when to escalate.

Think in cost per task

Because it is a verbose reasoning model, comparing M3.1 Flash on input price alone is misleading. A useful comparison includes input price, output price, cache discounts (90% here), how many output tokens a typical task produces, how often it needs a retry, and whether it completes the task at all. That is what cost per completed task measures, and Estimate AI Agent Cost shows how to work it out. Third-party API resellers also charge more than MiniMax’s own list price, so check which endpoint you are actually using.

Limitations

• Verbose output raises real cost and latency compared with what the token price suggests.
• Not the strongest model for hard reasoning or judgment-heavy work.
• Custom license with attribution and a revenue threshold, not Apache or MIT.
• Data location. If you use MiniMax’s hosted API, check where requests are processed and how they are retained. Self-hosting the weights avoids that question.
• Very new. Independent coverage beyond Artificial Analysis was still limited in its first week.

Alternatives

• DeepSeek V4.1 Flash, an MIT-licensed open-weight execution model with peak and off-peak pricing
• Mercury 3, a much faster diffusion model in the low-cost tier, API only
• Qwen3.8-27B, a smaller open model for more modest hardware
• Muse Glimmer, Meta’s 30B open model that runs on one consumer GPU
• OpenAI Decisions API, when the worker’s job is classification or routing rather than generating text

See also How to Make Cheaper AI Models Work Smarter and the Qwen vs Kimi vs DeepSeek vs GLM comparison.

Compare: MiniMax M3.1 Flash vs DeepSeek V4.1 Flash.