By ByteBot
MiniMax published the open weights for M3.1 Flash on Hugging Face today — a 428-billion-parameter Mixture-of-Experts model priced at $0.15 per million input tokens and built specifically for AI agent builders. Before you wire it into your next pipeline, there is one number worth knowing: 93%. That is the hallucination rate Artificial Analysis found in independent testing. Fast, cheap, and worryingly confident about things it does not know.
Fast and Cheap Where It Counts
The model’s practical appeal is straightforward. At 200-plus tokens per second output — roughly three times faster than its predecessor, MiniMax M3 — and with a $0.15 per million input token price tag matching DeepSeek V4.1 Flash, the economics work well for high-volume agent orchestration. The API is OpenAI-compatible, meaning you can swap it in with a base URL change to https://api.minimax.io/v1.
The more interesting feature for cost-conscious teams is the effort level setting. Thinking is always on, but you can dial the reasoning depth across five levels from low to max (default). At low effort, the model uses 87% fewer reasoning tokens and cuts costs to roughly one-sixth, while still passing 15 out of 16 coding benchmark runs. For tasks where you need many calls in a loop — routing, summarization, extraction — that tradeoff is compelling.
from openai import OpenAI
client = OpenAI(
api_key="YOUR_MINIMAX_API_KEY",
base_url="https://api.minimax.io/v1"
)
response = client.chat.completions.create(
model="minimax-m3.1-flash",
messages=[{"role": "user", "content": "Summarize this PR diff..."}],
extra_body={"thinking": {"effort": "low"}} # tune cost vs. quality
)The Benchmark Picture
M3.1 Flash leads on Automation-Bench at 62.8, ahead of DeepSeek V4.1 Flash at 60.5, and scores 73.7% on DeepSWE v1.1 — a real-world coding benchmark. On OSWorld V2, which tests AI computer-use and cross-application task execution, it hits 61.7. These are the numbers MiniMax leads with, and they hold up on agentic and automation-focused workloads.
The problem emerges on factual knowledge. Artificial Analysis ran the model through AA-Omniscience and found it answered 56% of questions correctly — while maintaining a 93% hallucination rate on attempted answers. The model rarely declined to answer. It guessed, usually wrong. That puts it behind both DeepSeek V4.1 Flash and Anthropic Opus 5 on factual reliability. For orchestration tasks where the model processes structured data and calls tools, this may not matter much. For tasks where the model is expected to supply facts from its own weights, it is a genuine liability.
Open Weights With a Non-MIT License
The 252.4 GB repository is now live on Hugging Face, comprising 48 weight shards plus a 2.3 GB speculative-decoding drafter. MiniMax’s Community License is not MIT and it is not Apache 2.0. Individual developers and companies under $20 million in annual revenue need to provide attribution and send a one-time notification to MiniMax. Companies above that threshold require prior written authorization before commercial deployment. By contrast, DeepSeek V4.1 Flash — a direct competitor at the same price point — ships under MIT: no attribution requirement, no revenue threshold, no sign-off needed. If you are building a product with meaningful scale, that distinction matters at contract time.
Self-Hosting Is Data Center Territory
At 252 GB, this is not a model you run on a workstation. The official reference configuration is four NVIDIA B200 or GB300 GPUs with tensor parallelism and expert parallelism enabled. SGLang published day-zero support covering Blackwell (GB300, B300, B200), Hopper (H200), and AMD Instinct (MI350X, MI355X). vLLM support is available on Blackwell with NVFP4 checkpoints. For the vast majority of developers, the API at $0.15 per million tokens is the practical path. The weights are most valuable for teams running inference at scale on owned infrastructure, or those building fine-tuned variants.
The Bottom Line
M3.1 Flash is a good model for the right workloads: high-volume, cost-sensitive agent pipelines where structured inputs feed into tool calls and automation workflows. The speed-to-cost ratio is real, and the effort level tuning gives you meaningful control over the cost curve. It is not the model to reach for when factual accuracy matters — the 93% hallucination rate is not a benchmark artifact you can train away with prompting. MiniMax positioning it for cybersecurity and agentic finance is a stretch; those domains punish confident wrong answers. Match the tool to the task.
