I have a rule for new model releases: I care less about the press release than about what a model costs to run and whether it survives contact with a real codebase. Most releases fail one of those two tests. MiniMax M3.1 Flash, announced October 19, 2026 with weights released October 23, passes both — and that is why it deserves your attention this week.
I build websites and tools for a living, and my coding assistant is one of my biggest monthly bills after hosting. When a model shows up that is faster, cheaper per task, and open-weight on top of that, it changes what I can afford to automate. M3.1 Flash is exactly that kind of release.
What MiniMax actually shipped
Let me untangle the timeline, because this release came in two parts and people keep mixing them up. On October 19, MiniMax launched M3.1 Flash as a hosted API service at platform.minimax.io, alongside its MiniMax Code desktop app. It had already previewed anonymously as "Space Bunny" before the announcement.
Then, on October 23, MiniMax released the model's weights, as it had promised. That second step is the one that matters most for developers who like to own their stack. The weights release means you can download the model and run it yourself, rather than renting it through someone else's API forever. TechNode reported the release on the same day, and deployment guidance came with it for the inference frameworks developers actually use.
So remember there are two products here: a cheap hosted API and an open-weight model you can self-host. Most models give you one or the other; getting both at once is why this release travelled so fast this week.
The specs that matter for working developers
M3.1 Flash is a mixture-of-experts model with roughly 428 billion total parameters, of which about 23 billion are active for any given token. That is the design trick behind the whole release: a huge model that only wakes up a small slice of itself per token, so it runs far cheaper than its headline size suggests.
The context window is 1 million tokens, with up to 512,000 tokens of output. In practical terms, that means you can hand it a large repository, a long specification, or a multi-step agent session and it will still have room to think. It accepts text, image, and video input, and it supports function calling, structured output, and web search — the full toolkit an agent needs to do multi-step work rather than just answer questions.
Thinking is always on with this model, but the depth is tunable, from low to maximum reasoning effort. For a quick bug fix you do not want to pay for deep reasoning; for a refactor across a dozen files, you do. A model that lets you turn that dial respects your budget.
On price, MiniMax charges $0.15 per million input tokens and $0.60 per million output tokens, with cached input at about $0.015 per million. Because the model also uses fewer output tokens than its rivals for the same work, the cost per task lands at roughly $0.22, according to Artificial Analysis figures cited across this week's coverage. MiniMax says the model runs about three times faster than the older M3, and its product page lists around 150 tokens per second. I treat vendor speed claims with suspicion until I measure them myself, but the direction is consistent: this is a model built to be fast and cheap, not just smart.
Benchmarks, read honestly
Benchmarks are marketing with numbers attached, so note who produced each number below.
The headline independent score comes from the Artificial Analysis Intelligence Index, where M3.1 Flash scores 42 points — 13 points higher than its predecessor M3, tying GLM-5.3-Flash, edging past Gemini 3.8 Flash at 41, and trailing DeepSeek V4.1 Flash at 45. That is a genuine jump in one generation, and it puts a Chinese open-weight model within touching distance of the best flash-tier models in the world.
On specific tests, the picture stays strong. On Humanity's Last Exam, a hard general-knowledge benchmark, M3.1 Flash hits 48 percent, beating DeepSeek V4.1 Flash at 43 percent and GLM-5.3-Flash. On CritPt, a research-level physics benchmark of 71 complex problems, it scores 22 percent — ahead of GLM at 15 percent and Gemini 3.8 Flash at 18 percent, just behind DeepSeek at 24 percent. On the AA-Omniscience factual-accuracy test, it answers 56 percent correctly, well above GLM at 28 percent and DeepSeek at 43 percent.
For coding specifically, MiniMax's own table puts M3.1 Flash at 73.7 percent on DeepSWE v1.1 at under $0.30 per task, essentially matching DeepSeek V4.1 Flash at 74.2 percent while costing less, plus 62.2 percent on Automation-Bench, 38.1 percent on Terminal-Bench, and 91.0 percent on BabyVision with tools. Those come from MiniMax, so treat them as the vendor's best case — but the independent numbers point the same way.
One number deserves remembering alongside all of that: a 93 percent hallucination rate on one Artificial Analysis measure, because the model tries to answer every test question instead of declining when it is uncertain. That is not a reason to dismiss it, but it is a reason to verify its output the way you would verify a confident junior developer's work. In coding, a wrong answer that looks right is the most expensive kind.
What "open weights" really buys you
Open weights get thrown around loosely, so let me be precise about what this release gives you and what it does not.
You can download the model and deploy it with SGLang or vLLM, the two inference frameworks most self-hosters already know, and MiniMax's release includes support for multiple chip platforms through FlagOS. AnchorTerminal's review notes the model needs a 4 to 8 GPU server for its roughly 250 GB of NVFP4 and MXFP8 weights. That is a small-server project, not a laptop one, but entirely realistic for a team or agency.
What open weights do not give you is a blank cheque. Commercial use carries attribution and notice requirements under the model's licence terms, so read the licence before you ship anything built on it. And note a wrinkle in the record: at least one platform's metadata has listed the weights as closed, even as MiniMax describes the model as open-weight. The weights release itself is confirmed, but the licensing fine print deserves a careful read rather than an assumption.
Still, the practical value is real: your code never leaves your server, your costs become hardware you already own, and no provider can change the price or retire the model under you. For agencies handling client code, that privacy argument alone is worth more than a few benchmark points.
How to try it today
The fastest path is the hosted API at platform.minimax.io, with the model ID MiniMax-M3.1-Flash. It works through OpenAI-compatible and Anthropic-compatible APIs, so most coding tools can point at it with a base-URL change rather than a rewrite. MiniMax also ships a desktop app called MiniMax Code if you want a ready-made coding assistant.
It is also available through aggregators. Vercel's AI Gateway carried it free for a short launch window until October 22, and it remains available through the gateway and other providers at standard rates. If you use Claude Code, Cursor, Codex, or a similar agent, you can route it through a gateway rather than wiring up a new provider directly.
My suggestion: start with the API on a real task you already understand, something with a known-good answer, so you can judge the output honestly. If the quality holds up, then look at self-hosting. The API is the cheap experiment; the server is the commitment.
Why this matters beyond one model
Zoom out: this is the latest move in a price war among Chinese labs that has no floor in sight. DeepSeek, GLM, Kimi, and now MiniMax all ship strong open-weight models at prices that undercut American frontier labs by multiples. When a model this capable costs $0.22 a task, the question stops being which model is smartest and becomes which is cheapest per finished job.
That shift favours developers and small teams more than anyone. A year ago, serious agent coding meant paying frontier prices or accepting a visibly weaker model. Now the middle of the market — fast, cheap, good enough for most real work — is crowded with open options. If you run a WordPress agency, a freelance practice, or a small product team, your agent-coding bill just became negotiable.
It also changes the self-hosting maths: renting is still easier, but owning is no longer a sacrifice in quality. Expect more teams to run a hybrid — hosted APIs for bursts, self-hosted models for steady, private workloads.
The catches I would not ignore
I have covered the good news, so here is the honest list of caveats. First, the hallucination rate: verify outputs, especially on factual questions, and never let an agent commit code you have not reviewed. Second, the licence: read the attribution and notice terms before commercial use. Third, the hardware: self-hosting needs a serious GPU server, so budget for that before promising a client a private deployment. Fourth, the ecosystem is young: vLLM and SGLang support was still pre-release in some reports as of October 24, so expect rough edges if you deploy in the first weeks.
None of these are deal-breakers — just the normal cost of adopting a model in its first week.
Conclusion
MiniMax M3.1 Flash is the rare release that is simultaneously a better model, a cheaper model, and a more open model than its predecessor. It scores 42 on the Artificial Analysis Intelligence Index, handles a million tokens of context, costs about $0.22 per task through its API, and its weights are out in the open for teams that want to run it themselves.
If you write code for a living, try it this week on a task you can judge: compare its output against whatever you use today and its price against your current bill. The open-weight price war is not slowing down, and the developers who benefit most will be the ones who test each new arrival instead of just reading about it.
FAQ
What is MiniMax M3.1 Flash?
It is MiniMax's fast, cost-efficient frontier model for coding, agentic work, computer use, and cybersecurity. It is a mixture-of-experts model with about 428 billion total parameters and 23 billion active per token, a 1-million-token context window, and text, image, and video input.
When was M3.1 Flash released?
MiniMax announced it on October 19, 2026 as a hosted API, after an anonymous preview under the name "Space Bunny." The open weights were released on October 23, 2026.
How much does M3.1 Flash cost?
Through MiniMax's API it costs $0.15 per million input tokens and $0.60 per million output tokens, with cached input at about $0.015 per million. Because it uses fewer output tokens per task, the cost per task works out to roughly $0.22.
Can I self-host M3.1 Flash?
Yes. The weights are open and deployable with SGLang or vLLM, with multi-chip support through FlagOS. Expect to need a 4 to 8 GPU server for the roughly 250 GB of weights, and check the licence's attribution and notice terms before commercial use.
How does M3.1 Flash compare to DeepSeek, Gemini, and GLM?
On the Artificial Analysis Intelligence Index it scores 42, tying GLM-5.3-Flash, edging past Gemini 3.8 Flash at 41, and trailing DeepSeek V4.1 Flash at 45. On Humanity's Last Exam it scores 48 percent, ahead of DeepSeek's 43 percent.
What should I watch out for with M3.1 Flash?
Its hallucination rate: Artificial Analysis measured about 93 percent on one test because the model answers every question instead of declining when unsure. Verify its output, review any code before committing it, and read the licence terms before commercial deployment.
