watch Minimax M3 1 Flash · Research

MiniMax M3.1 Flash: 1M-Context Coding Model That Always Thinks

Data graphic: MiniMax M3.1 Flash, a multimodal coding model that always thinks before it answers, shown with a huge "1M" for its 1-million-token context window. Two smaller stats give its 5 effort levels low to max, default max and a 66.25% score 53/80 points on KingBench 3, an independent 8-task test.
AK

Threat intelligence editor · Published Oct 2, 2026 · Updated Oct 24, 2026, 6:18 PM EDT

MiniMax's M3.1 Flash preview brings a 1M-token context, image and video input and five effort levels to MiniMax Code. What it does, and how it tests.

MiniMax has launched MiniMax-M3.1-Flash-Preview, a coding model with a 1-million-token context window. It went live on 27 September 2026 inside MiniMax Code, the company's agentic coding tool. MiniMax describes it as a "frontier multimodal coding model" built for agentic reasoning, tool use, coding and long-context work. It is the first M3.1 model to ship.

Update, 25 October 2026: open weights, a model card and an API price

MiniMax has dropped the "Preview" tag and released MiniMax-M3.1-Flash as an open-weight model. Its Hugging Face repository was created on 17 October 2026, and the weights landed in a single commit dated 23 October. MiniMax's API docs now list the model as MiniMax-M3.1-Flash, and its pay-as-you-go price list charges $0.15 per million input tokens, $0.60 per million output tokens and $0.015 per million cached input tokens. OpenRouter added it on 19 October at the same input and output price.

The model. The card describes a native multimodal mixture-of-experts model with about 428B total parameters and 23B active per token, the same 1M-token context and text, image and video input. M3.1 Flash keeps M3's architecture with three changes: sparse attention in every layer, quantization-aware training of the routed experts and attention paths, and a DSpark speculative-decoding draft head shipped with the weights.

The parameter count. Hugging Face's own metadata for the repository reports about 220B parameters, which does not match the card. Most of the gap comes from how the weights are stored. The config shows the routed experts are packed in NVFP4, two 4-bit values to a byte, and Hugging Face counts each of those 206.6 billion packed bytes as one parameter. Counting two weights per byte gives about 427B, close to the card's figure. That is our arithmetic, not a number MiniMax publishes. The full download is about 252 GB.

The licence. The weights use the MiniMax Community License, not an open-source licence. It grants free use for non-commercial purposes. Commercial use is allowed, but you must show "Built with MiniMax M3.1-Flash" on the product and send MiniMax a one-time notice, and products with more than $20 million a year in revenue need MiniMax's prior written authorisation. Fine-tuned derivatives offered commercially count as commercial use. The licence also bans uses including any military purpose.

The benchmarks. These scores are vendor-reported on MiniMax's model card, and we have not reproduced them. MiniMax ran most of the coding tests in agent harnesses such as Claude Code and Codex. A dash means the card gives no score.

Benchmark (vendor-reported)M3.1 FlashDeepSeek V4.1 Flash-1017Kimi K3GLM-5.3
CyberGym82.2—80.084.5
DeepSWE v1.173.777.067.566.9
Terminal Bench 4.038.142.312.637.9
NL2Repo-Bench61.268.258.058.0
Automation-Bench v1.0.662.860.546.748.8
BankerToolBench76.2—69.174.5
OSWorld 2.061.7—58.3—
LOCA-Bench 256K60.3—49.750.3

On MiniMax's own numbers, M3.1 Flash beats Kimi K3 on every row above, but GLM-5.3 and DeepSeek V4.1 Flash-1017 each lead it somewhere. On CyberGym, the cybersecurity test MiniMax highlights, GLM-5.3 leads at 84.5, M3.1 Flash scores 82.2 and Kimi K3 80.0, while the card gives no DeepSeek V4.1 Flash-1017 score. DeepSeek V4.1 Flash-1017 leads M3.1 Flash on DeepSWE, Terminal Bench and NL2Repo. The same table also lists GPT-5.6 Sol and Opus 5, and Opus 5 has the top score on most of the coding rows.

What fits on your own hardware. Community GGUF and MLX builds followed from 23 October. File sizes from their Hugging Face listings:

  • 3-bit and up does not fit a 128 GB machine. Unsloth's smallest GGUF, the 3-bit UD-IQ3_S, is 175.5 GB, and its UD-Q3_K_XL is 194.0 GB. The mlx-community 3-bit MLX build is 186.5 GB. These need a 192 GB or 256 GB machine once you leave room for context.
  • Under 3 bits, a 128 GB machine is borderline. AtomicChat's 2-bit AD-IQ2_XXS is 112.2 GB and loads in stock llama.cpp; its card says it fits a 128 GB Mac only with the GPU memory limit raised and context up to about 32k tokens. Smaller experimental 1-bit files (94.8 GB and 106.6 GB) need a patched llama.cpp or a custom runtime, and their uploader measures large quality losses at that size.
  • 4-bit is around 250 GB. Unsloth's UD-Q4_K_XL is 249.4 GB, barely smaller than the 252 GB official release, which MiniMax recommends serving on GPUs with SGLang, vLLM or TokenSpeed.

What M3.1 Flash is

According to MiniMax's API documentation, M3.1 Flash:

  • Holds 1,000,000 tokens of context, which MiniMax pitches at long documents, whole codebases and multi-step agent sessions.
  • Takes text, image and video input, so screenshots, diagrams and screen recordings can go into the same prompt as the code.
  • Always thinks before it answers. Reasoning cannot be switched off: a request that sends thinking: {"type": "disabled"} or effort: "none" is rejected with HTTP 400 and the message that the model "requires adaptive thinking".
  • Has five effort levels: low, medium, high, xhigh and max. Higher levels think longer, produce more output tokens and take more time. If a request leaves effort out, the model runs at max.

That last default matters in practice. A tool that does not set effort gets the slowest, most token-hungry setting on every call. For quick edits and autocomplete-style tasks, set low or medium explicitly.

Data graphic: MiniMax M3.1 Flash always thinks. A request passes through one of five effort levels low, medium, high, xhigh, max into adaptive thinking that can't be turned off, then to the answer. Leaving effort unset runs it at max, and disabling thinking or sending effort none returns an HTTP 400 error.

Every M3.1 Flash request runs through always-on adaptive thinking: leaving effort unset means max, and turning thinking off returns HTTP 400.

How it performs

When this article first ran, MiniMax had published no benchmark scores, model card, parameter count or open weights for M3.1 Flash. It has since released all four (see the update above), though it still gives no speed figure.

The clearest independent result so far comes from the AICodeKing channel's KingBench 3, a set of eight build-it-from-scratch coding tasks scored out of 10 each. Tested on 28 September, M3.1 Flash scored 53 of 80 (66.25%). The best result in the same run was Claude Opus 5.5 at 93.75%.

The scores were uneven:

  • Strong on 3D geometry. It scored 9/10 on a folding-table task, with smooth animated 3D geometry.
  • Weak on interactive simulations. An elevator simulation crashed on load because the code called a position value as a function (3/10). An archery game drew its targets in the wrong place and paused its timer between shots (4/10).

Some sites also circulated figures such as a 73.8% SWE-bench Verified score, 165 tokens per second and a $0.10 per million input-token price. None of them matches what MiniMax has since published. Its model card gives no SWE-bench Verified score or speed figure, and its price list charges $0.15 per million input tokens, not $0.10.

How to get it

At launch, the preview was "available only through M Plan and MiniMax Code for now". That has changed. There are now three ways in:

  • MiniMax Code, on the desktop app (macOS or Windows) or the web, where the preview launched.
  • The pay-as-you-go API, as model MiniMax-M3.1-Flash, at $0.15 per million input tokens and $0.60 per million output tokens. OpenRouter also lists it, so you can put it behind a gateway alongside the models you already use.
  • The open weights on Hugging Face, which MiniMax recommends serving with SGLang, vLLM or TokenSpeed.

Whichever route you take, set the effort level per task rather than leaving it on the default max.

From 1 to 7 October 2026, MiniMax gave subscribers unlimited use of the preview in MiniMax Code. That offer has ended.

Data graphic: In AICodeKing's independent KingBench 3 test, MiniMax M3.1 Flash scored 9/10 on a 3D folding-table task but only 4/10 on an archery game and 3/10 on an elevator simulation. Its total was 53/80 66.25% across all 8 tasks, an average of 6.6/10 per task.

M3.1 Flash scored 9/10 on a 3D geometry task but 4/10 and 3/10 on two interactive builds, for 53/80 (66.25%) overall on AICodeKing's independent KingBench 3.

Should you use it?

M3.1 Flash is worth testing now on big-context work, such as reading a large repository or reviewing long agent sessions, and on visual tasks that need image or video input. It now has a per-token price and MiniMax's own benchmarks, but those benchmarks are vendor-reported and the independent test above is uneven, so:

  • Run it side by side with your current model on your own tasks before you make it the default in a production pipeline.
  • Check its output on anything interactive or stateful, where the independent tests show it is weakest.
  • Read the licence before you ship it in a product: commercial use needs attribution and a notice to MiniMax, and authorisation above $20 million in yearly revenue.
  • Don't plan on running it locally on a 128 GB machine. Only sub-3-bit builds fit, and only just.

Sources

Keep reading

All latest →
  1. watchResearchGitHub Enterprise Now Includes Copilot at $0, but Chat, Code Review and the Cloud Agent Bill From the First Request3 min
  2. watchResearchChatGPT Pro and Business Now Include Up to $500 a Month in API Credits: Who Actually Gains4 min
  3. watchResearchNemotron 3.5 Transcribe Posts 4.71% English WER in Open Weights, and Fits a 32 GB GPU Only in 4-Bit4 min
  4. watchResearchLFM3-1.7B vs Qwen3.5-9B: The 1.7B Model Wins on Tool Calls, Then Loses on Reasoning and Long Context3 min
  5. stableResearchOllama 0.40.11 to 0.40.13: An OpenAI-Compatible Speech Endpoint, d3 Decision Models, LFM3 and DFlash24 min
  6. watchResearchNanbeige4.3-33B vs Qwen3.8-27B: Where the Vendor Benchmarks Diverge, and Which GGUF Fits 16, 24 or 32 GB5 min