MiniMax published the weights for MiniMax-M3.1-Flash on October 23, four days after the model's API launch, turning a proprietary Flash-tier release into something developers can download and run themselves.
The Hugging Face repository lists an image-text-to-text pipeline with safetensors weights loadable through transformers, tagged multimodal, moe, agent and coding. Together AI is providing hosted production inference alongside the release, giving teams a route that does not require managing GPUs.
What the model is
M3.1 Flash is a sparse mixture-of-experts model with roughly 428 billion total parameters and about 23 billion active, the same shape as M3. It carries a one-million-token context window, a native visual encoder, and MiniMax Sparse Attention in place of full attention — KV-block selection that cuts per-token compute at long context and, according to the company, brings roughly one-twentieth the cost of the previous generation at a million tokens.
Inputs are text, image and video; output is text. Reasoning is always enabled and cannot be switched off, with effort selectable across low, medium, high, xhigh and max. Reported throughput is around 150 tokens per second, though that is an input/output figure rather than an end-to-end task latency guarantee.
MiniMax frames the model for agentic co-work, codebase comprehension and repair, computer use, and cybersecurity — including code auditing, vulnerability patching, penetration testing, malware analysis and reverse engineering. The company also disclosed that "Space Bunny," a model that operated anonymously and processed more than 200 trillion tokens, was an early version of this release.
The numbers, and who produced them
MiniMax's own figures put the model at 61.7 on OSWorld V2, which measures cross-application computer operation, and 29.6 on Agents' Last Exam, which tests economically valuable professional work. It says the model ranks first among Chinese models on both. It also claims inference roughly three times faster and overall usage cost 70 percent lower than its predecessor.
Independent evaluation is more sober. On Artificial Analysis, M3.1 Flash scores 42.7 on the Intelligence Index, placing 21st of 351 models — one place above GPT-5.6 Terra at 42.1. Its Humanity's Last Exam result is 48.0 percent, AA-LCR 79.7 percent, SciCode 53.4 percent, and Terminal-Bench 4.0 just 25.8 percent, 22nd of 124. The AA-Omniscience non-hallucination rate is 7.2 percent, a figure worth reading twice before trusting the model's factual output without verification.
The gap between MiniMax's benchmark table and the independent one is not evidence of dishonesty. It is what happens when a vendor evaluates on its own harness and a third party evaluates on another, and it is the reason neither number should be quoted alone.
The licence is the story
Open weights are not open source, and MiniMax's licence makes the distinction concrete. M3.1-Flash ships under the MiniMax Community License, not MIT. Non-commercial use is free. Commercial use must display "Built with MiniMax M3.1-Flash." Organizations with annual revenue above $20 million require prior written authorization.
That is a meaningfully different proposition from DeepSeek's MIT-licensed V4.1-Flash-1017, released a week earlier at the same $0.15 input and $0.60 output per million tokens. Two Chinese labs, two near-identical price points, two incompatible legal positions on what an enterprise may do with the weights.
Pricing and access
List pricing is $0.15 per million input tokens, $0.60 per million output, and $0.015 per million cached input — flat, with no peak/off-peak split, unlike DeepSeek. The free launch period on OpenRouter ended October 22 at 12:00 UTC, and the :free variant has been retired; the paid identifier is the one to use.
Why it lands now
The release is best understood as a distribution decision rather than a research one. Weights arriving ahead of supporting documentation is a pattern among Chinese model builders: the Hugging Face repository functions as the first public announcement.
For MiniMax the calculation is competitive. DeepSeek has spent the month making itself the default for cost-sensitive agentic work — MIT weights, a million-token window, and a benchmark table that beats Kimi K3 on ten of eleven tests. MiniMax cannot out-open DeepSeek on licence, so it is competing on throughput, on multimodal input, and on the claim that its cost per completed task, rather than per token, is lower.
Whether that holds is not yet established. Long tool-calling chains and long-context stability are the capabilities enterprise buyers actually test, and neither has published success rates. Until independent reproductions appear on agent benchmarks, the weights are public and the evidence is not.