What is MiniMax-M3.1-Flash?
MiniMax-M3.1-Flash is MiniMax's fast reasoning model, released on 19 October 2026 after a preview inside MiniMax Code. MiniMax says an earlier version ran anonymously as "Space Bunny" and processed more than 200 trillion tokens. Weights followed on Hugging Face on 23 October under a MiniMax community licence, the same terms as MiniMax-M3.
Despite the name it is not a small model. It keeps M3's mixture-of-experts design, about 428 billion parameters with 23 billion active per token, and reads text, images and video into a 1M-token context. MiniMax replaced the first three full-attention layers in M3 with sparse attention, trained the experts with quantisation in the loop and added a DSpark speculative-decoding head; the published checkpoint is about 250 GB in MXFP8.
Reasoning cannot be switched off. Effort runs from low to max in five steps, and max is the default. MiniMax's API charges $0.15 per million input tokens, $0.60 per million output and $0.015 for cached input at every prompt length.
Is MiniMax-M3.1-Flash worth it?
M3.1-Flash is a large step up from M3 at half the price, and its weakness is knowing when it does not know. Artificial Analysis scores it 43 on its Intelligence Index, fifth of 173 models in its class against a median of 13, where M3 scored 29. Its cost per index task is $0.22, against about $0.51 for M3. VentureBeat, reading the same firm's AA-Omniscience results, reports that it answered 56% of questions correctly, ahead of DeepSeek's 43%, but hallucinated in 93% of the cases where it was not right, rather than declining or hedging.
MiniMax's own card, all self-reported, puts it ahead of DeepSeek-V4.1-Flash-1017 on Automation-Bench v1.0.6 (62.8 to 60.5) and level on Humanity's Last Exam without tools (43.4 to 42.6). On coding it trails: 73.7% to 77.0% on DeepSWE v1.1, 38.1 to 42.3 on Terminal Bench 4.0 and 61.2 to 68.2 on NL2Repo-Bench. Anthropic's Opus 5 leads both on most of the table. The card also ties Opus 5 at 76.2 on BankerToolBench, though that run was graded by MiniMax's own M2.7 model.
The always-on reasoning sets the real bill. The default is max effort, which MiniMax's docs say spends more output tokens and adds latency, and Artificial Analysis calls the model very verbose: 180 million output tokens to run its index, against a median of 100 million. Dropping effort per request is the lever MiniMax recommends, since sending a request with thinking disabled returns an error. Speed claims also need discounting. MiniMax says median output exceeds 170 tokens a second; Artificial Analysis measured 116.2 on MiniMax's API, around the class median, with 1.10 seconds to first token.
The weights shipped on time, under the licence that already slows M3 down: a "Built with MiniMax M3.1-Flash" credit and a notice email for any commercial use, prior written permission above $20 million in annual revenue, and no military use. At about 250 GB the checkpoint is a third the size of M3's BF16 release and fits on one eight-GPU H100 node, which makes M3.1-Flash the easier of MiniMax's two big models to bring in-house.
Where it earns its place
An operations team automating back-office work across a dozen SaaS tools
Run multi-step automations that read screenshots and fill in forms
MiniMax reports 62.8 on Automation-Bench v1.0.6 and 61.7 on OSWorld 2.0, both self-reported. At $0.15 input and $0.60 output, 10 billion input and 1 billion output tokens a month comes to about $2,100 before cache hits, and setting effort to low or medium on routine steps cuts the reasoning tokens billed as output.
A legal-tech engineer working on M&A data rooms
Ask questions across a whole contract bundle in one prompt
Price does not rise past 512K tokens as M3's does, so a full 1M-token prompt costs $0.15 in input, a quarter of M3's long-context rate, and $0.015 on cached re-reads. Given the hallucination rate, every answer should quote the clause it relies on so a reviewer can check it.
A support lead at a 200-person software company
Draft replies to customer tickets from the help centre
1M replies a month at 2,000 input and 300 visible output tokens is about $480, plus the reasoning tokens that the default max effort adds on top. With a 93% hallucination rate on questions it gets wrong, drafts need retrieval from the help centre and a human send button rather than auto-replies.
Before you commit
- Setup
- API: create a key on platform.minimax.io and call MiniMax-M3.1-Flash through the Anthropic-compatible endpoint (recommended) or the OpenAI-compatible one, setting reasoning_effort per request. Weights: hf download MiniMaxAI/MiniMax-M3.1-Flash and serve with SGLang, vLLM or TokenSpeed; MiniMax recommends temperature 1.0 and top_p 0.95.
- Integrations
- MiniMax API, MiniMax Code and the M Plan subscription; on OpenRouter at the same $0.15/$0.60 from MiniMax and about a dozen hosts including Together, Nebius and Novita. Transformers, KTransformers and Unsloth support is listed as in progress.
- Data and privacy
- Licence: commercial use needs a "Built with MiniMax M3.1-Flash" credit and a one-time notice to MiniMax; products over $20 million in yearly revenue need prior written authorisation; military use is prohibited. MiniMax's API retention and training terms were not published on the pages we could read.
- Cost at scale
- 10 billion input and 1 billion output tokens a month is about $2,100 on MiniMax's API, with cached input at $0.015 per million; reasoning at max effort inflates the output side. Self-hosting needs about 250 GB of GPU memory for the MXFP8 weights before KV cache, within an eight-GPU H100 80 GB node.
- Limits
- 1M-token context; OpenRouter lists 524,288 output tokens on MiniMax's own endpoint. Thinking cannot be disabled, and a request that tries returns a 400 error. MiniMax publishes no rate limits for the model, and Artificial Analysis's listing still described the weights as unreleased when we checked.
What's good
- $0.15 / $0.60 per 1M tokens at any prompt length
- Intelligence Index 43, fifth of 173 in its class
- Text, image and video input with 1M context
- About 250 GB checkpoint fits one 8x H100 node
What's not
- 93% hallucination rate on AA-Omniscience, per VentureBeat
- Reasoning cannot be disabled; default max effort is verbose
- Licence needs attribution, a notice and permission above $20M revenue
- Trails DeepSeek-V4.1-Flash-1017 on coding benchmarks (self-reported)
Best for
Teams running high-volume agent and automation work, or long multimodal inputs, who can verify outputs and accept MiniMax's licence terms.
MiniMax-M3.1-Flash alternatives
Head to head
MiniMax-M3.1-Flash: common questions
- Is MiniMax-M3.1-Flash worth it?
- Near the top of the open-weight table for a fraction of a cent per task, as long as nothing it says goes out unchecked. We rate it 4.0 out of 5.
- How much does MiniMax-M3.1-Flash cost?
- Open weights under the MiniMax community licence. MiniMax API per 1M tokens, flat at any prompt length: $0.15 input, $0.60 output, $0.015 cached input.
- What are the downsides of MiniMax-M3.1-Flash?
- 93% hallucination rate on AA-Omniscience, per VentureBeat; reasoning cannot be disabled; default max effort is verbose; licence needs attribution, a notice and permission above $20M revenue; trails DeepSeek-V4.1-Flash-1017 on coding benchmarks (self-reported).
- Who is MiniMax-M3.1-Flash best for?
- Teams running high-volume agent and automation work, or long multimodal inputs, who can verify outputs and accept MiniMax's licence terms.
- How much does MiniMax-M3.1-Flash cost for a team or in production?
- 10 billion input and 1 billion output tokens a month is about $2,100 on MiniMax's API, with cached input at $0.015 per million; reasoning at max effort inflates the output side. Self-hosting needs about 250 GB of GPU memory for the MXFP8 weights before KV cache, within an eight-GPU H100 80 GB node.
- Is MiniMax-M3.1-Flash safe for company data?
- Licence: commercial use needs a "Built with MiniMax M3.1-Flash" credit and a one-time notice to MiniMax; products over $20 million in yearly revenue need prior written authorisation; military use is prohibited. MiniMax's API retention and training terms were not published on the pages we could read.
- What are the best MiniMax-M3.1-Flash alternatives?
- The closest alternatives we have reviewed are DeepSeek-V4.1-Flash, MiniMax-M3, GLM-5.3 and Kimi K3.
Answers checked Oct 26, 2026.
Make MiniMax-M3.1-Flash? Show this rating on your site
