MiniMax M3.1-Flash-Preview Review 2026: 1M Context, No Price Tag, and a Thinking Mode You Cannot Turn Off
MiniMax M3.1-Flash-Preview Review 2026: 1M Context, No Price Tag, and a Thinking Mode You Cannot Turn Off
Verdict first: MiniMax M3.1-Flash-Preview went live on September 27, 2026 and reached public beta the next day. It ships with a 1,000,000-token context window, accepts text, image and video, exposes five reasoning-effort levels, and costs — well, nobody will tell you. There is no published per-token price, no model card, no benchmark table, and no parameter count. It is reachable only through MiniMax Code and the Token Plan subscription. Two things genuinely stand out. First, thinking cannot be turned off: sending thinking: {"type": "disabled"} returns HTTP 400, and the default effort level when you omit the parameter is max, the slowest and most expensive setting. Second, a stack of independent tokenizer probes ties this model to Space Bunny Alpha, the anonymous free model that topped OpenRouter and OpenCode call charts in September — a link MiniMax has never confirmed. The launch did not land well: stock fell roughly 9 to 10 percent. This is a capable-looking spec sheet wrapped in an unusual amount of silence, and the silence is the story.
MiniMax M3.1-Flash-Preview — 1M context and native multimodal input, launched with no published price, no benchmarks and no way to switch thinking off
What MiniMax M3.1-Flash-Preview Is
On September 27, 2026, MiniMax added a model called MiniMax-M3.1-Flash-Preview to its own tooling. There was no announcement, no blog post, no model card and no pricing entry. It went live inside MiniMax Code and the Token Plan subscription, and public beta followed on September 28. It is the first new M-series model in 120 days — the previous generation, M3, arrived on June 1, 2026.
MiniMax describes it as a native multimodal, frontier coding model aimed at everyday development work: fixing small bugs, reviewing code, building small features, and verifying tests. That framing matters, because it positions the model for short, frequent tasks rather than long-horizon agent runs or flagship general reasoning — which makes the inability to disable thinking a sharper problem than it would otherwise be.
On the same day, MiniMax reset every user's Token Plan quota, and from September 28 through October 7 it doubled check-in points inside MiniMax Code. Those are the signals of a launch. Everything else — the spec sheet, the price, the capabilities — was either left blank or contradicted elsewhere.
The Specifications
What MiniMax did publish is a real, if partial, specification sheet:
| Specification | Value |
|---|---|
| Model ID | MiniMax-M3.1-Flash-Preview |
| First seen | September 27, 2026 (MiniMax Code) |
| Context window | 1,000,000 tokens (input + output combined) |
| Input modalities | Text, image, video |
| Output modality | Text only |
| Video limit | URL or base64, up to 50 MB |
| Image limit | Up to 10 MB |
| Request body limit | 64 MB |
| Reasoning effort | Five levels: low, medium, high, xhigh, max — defaults to max |
| Disable thinking | Not supported |
| Endpoints | Anthropic-compatible (recommended), OpenAI-compatible, OpenAI Responses |
| Maximum output | Not published |
| Output speed | Not published (one source claims 100+ tok/s; sources conflict) |
| Parameter count | Not published |
| Open weights | None |
Two practical notes. The model has no /models discovery capability, so you must write the model name by hand — you cannot enumerate what is available. And the reasoning_effort parameter, while central to how this model is meant to be controlled, is documented as effective only for M3.1-Flash-Preview; it is not a general M-series control you can carry across models.
Tool use is supported: function calling and structured output both work, so the model is usable as the reasoning step inside a larger pipeline. What it cannot do is stop reasoning, which brings us to the most consequential design decision in the release.
The Space Bunny Connection
Before M3.1-Flash-Preview appeared, something more interesting did. On September 23, 2026, a model called Space Bunny Alpha showed up on the aggregation platform OpenRouter and the coding platform OpenCode. It had a 1M-token context, accepted text, image and video, offered adjustable reasoning depth, and was completely free. During the Mid-Autumn Festival it topped the daily call-volume charts on both platforms. Nobody knew who made it.
The evidence linking it to MiniMax came from an unusual direction: tokenizer fingerprinting, essentially an identity stress test. Independent researchers pushed token-count probes through the Space Bunny endpoints and compared the results against known models on the same gateway.
| Tester | Result |
|---|---|
| Independent researcher (Sept 24) | 24 token-count test sets; Space Bunny matched MiniMax models on the same gateway more closely than Kimi or GLM |
| Second tester | Text and vision tokenizers matched MiniMax M3 exactly |
| Third tester | 12 text samples / 36 comparisons all matched MiniMax M3 — Qwen, GLM, DeepSeek and Llama did not |
| Fourth tester | 50 of 50 test strings matched |
| "stealthprint" | 24 of 24 probes matched |
The conclusion was consistent: Space Bunny shares a tokenizer with the MiniMax M3 family. Because Space Bunny carried the "Flash" label and matched the announced spec points, developers on Reddit concluded it was very likely the preview build of MiniMax M3.1 Flash.
Two caveats are essential here, and they are not small. First, MiniMax has never publicly claimed Space Bunny; as of publication the model remains officially anonymous. Second, a shared tokenizer is not the same as the same model — different models can share tokenizers, and gateways can process inputs and outputs before they are counted. So this is strong correlation, not proof.
The practice itself is becoming ordinary. Zhipu's GLM-5.3-Flash was blind-tested under the codename "Ox-Alpha" in August 2026. Vendors use these anonymous previews to gather real-world feedback and stress-test demand and server capacity before revealing who they are — which is exactly the kind of thing you would do if you wanted honest evaluations without the marketing baggage, or without committing publicly to a spec sheet.
The Thinking Mode You Cannot Turn Off
This is where developer reaction concentrated, and for good reason.
Every MiniMax M-series model reasons before it answers. M3.1-Flash-Preview is the first that refuses to stop. If you try to disable reasoning — the standard move when you want a fast, cheap response from a model that does not need to deliberate:
{"thinking": {"type": "disabled"}}
// or
{"reasoning_effort": "none"}the API returns HTTP 400 with error code 2013:
model "MiniMax-M3.1-Flash-Preview" requires adaptive thinking;
thinking.type="disabled" (including reasoning.effort=none)
is not allowed (2013)There is no supported way to obtain a non-reasoning response. MiniMax's own guidance is to stop trying and lower the effort instead. The five-level ladder is the intended latency and cost lever:
| Effort level | Intended use |
|---|---|
low | Simple syntax fixes, quick lookups |
medium | Routine refactors |
high | Complex logic |
xhigh | Hard debugging, architecture work |
max | High-risk or ambiguous problems — the default when omitted |
Three related limitations compound the issue. The reasoning_split switch, which separates thinking into a reasoning_content field so you can log or hide it, cannot currently be set to false on this model. The effort parameter applies only here, so you cannot reuse the same control pattern across the M series. And where the thinking text actually lands depends on which protocol you use: the Anthropic-compatible endpoint returns a thinking content block, the OpenAI-compatible endpoint returns it on reasoning_content, and the Responses endpoint returns a reasoning output item.
The practical consequence: if you want low latency you must manually dial effort down on every call, and the default is max — the slowest and most expensive tier. Omit the parameter once and both your bill and your wait time climb. For a model explicitly positioned at "everyday development" work like small bug fixes and code review, making maximum deliberation the default is a curious choice.
No Price Tag: Subscription Only
MiniMax published no per-token price for this model. It does not appear on the public pay-as-you-go rate card, which sets it apart from both M3 and M2.7 — every previous M-series model had a metered rate.
| Access method | Details |
|---|---|
| Token Plan | Subscription, tiers quoted at roughly $22 / $55 / $132 per month |
| MiniMax Code | First-party coding tool; uses a Token Plan subscription key, not a metered API key |
| Pay-as-you-go API | Not available |
Because there is no metered rate, cost can only be estimated as subscription divided by usage. That is workable for a team on a fixed budget — you cannot be surprised by a spike — but it is useless for unit economics. You cannot compute cost per request, cannot compare against a competitor on price, and cannot model the cost of a workload before committing to it.
There is also a documentation inconsistency worth flagging. MiniMax's Token Plan pricing footer still does not name M3.1-Flash-Preview, and continues to list M3 and M2.7. The model page says the model is available on Token Plan and nowhere else. Those two pages contradict each other, and the contradiction was still live after the public beta began.
The Benchmark Void
This is the most unusual part of the release, and it is worth stating plainly:
| Source | Status |
|---|---|
| Official benchmark table | None |
| Model card | None |
| Artificial Analysis model index | Not listed |
| BenchLM third-party board | Zero recorded results as of September 27 |
Normally a missing vendor benchmark is filled in by third parties within days. That cannot happen here, and the reason is structural: the model is only reachable through MiniMax Code, so outsiders cannot run standardized suites against it or measure throughput independently. The closed distribution that makes the model easy to monetize is the same thing that makes it impossible to evaluate.
This is the direct cost of a silent launch. You cannot verify any claim about the model because no claim has been made. If you are deciding whether to pay for a Token Plan to reach it, you are being asked to trust a spec sheet with the two most important rows — quality and cost — left blank.
How It Compares With MiniMax M3
The generational comparison is informative, with one large caveat: M3's numbers are vendor-reported, and M3.1-Flash-Preview has no numbers at all.
| Specification | M3.1-Flash-Preview | MiniMax M3 |
|---|---|---|
| Released | September 27, 2026 | June 1, 2026 |
| Context window | 1M tokens | 1M tokens |
| Pricing | Not published | $0.30 / $1.20 per million tokens |
| Speed | Not published | ~100 tok/s |
| SWE-bench Verified | Not published | 80.5% |
| SWE-bench Pro | Not published | 59.0% |
| Terminal-Bench 2.1 | Not published | 66.0% |
| Open weights | None | Yes (Hugging Face) |
| Access | MiniMax Code / Token Plan only | API + MiniMax Code |
The only confirmed additions over M3 are the five-level effort control and video input. On the metrics that actually determine whether you should use it — parameter count, price, speed, and task accuracy — the sheet is empty.
One item in the table deserves separate attention. M3 has open weights. M3.1-Flash-Preview does not. A restricted Hugging Face repository named MiniMax-M3.1-preview-private exists at roughly 250 GB, but outsiders cannot download it, and MiniMax has not responded to questions about whether it holds this model's weights. If it does, the model is effectively open-weight in storage and closed in practice.
Market Reaction: Down 9 to 10 Percent
A new model is normally a positive signal. This one was not received that way. MiniMax stock fell roughly 9 to 10 percent on the day, landing near HK$247.4 per share.
The drop reflects continuing investor concern about MiniMax's revenue structure, its ongoing losses, and the pressure to commercialize. A preview model with no published price and no clear path to scaled monetization does not address any of those concerns — and the Space Bunny speculation, however intriguing technically, did not move the stock either.
Unverified Claims
One leaked "partner architecture note" claims the model uses full sparse attention replacing M3's three full-attention anchors, Q8KV4 precision (FP8 query with 4-bit KV), W4A4 NVFP4 experts, and DSpark speculative decoding.
These claims come from an unidentified source, MiniMax has not commented, and their accuracy is unverified. They are recorded here as rumour only, and should not be treated as specification. Given that the vendor published no architecture at all, treat any third-party architecture claim about this model with the same caution.
The Honest Take
Who this is for: teams already inside MiniMax's ecosystem who hold a Token Plan, value the 1M context window for long-document or large-repository work, and want native video input. If you are already paying for Token Plan, M3.1-Flash-Preview is a free addition you can trial immediately, and the Anthropic-compatible endpoint makes it a low-effort swap into an existing pipeline.
Who should wait: anyone who needs to justify the cost, benchmark the model before adoption, run it on their own hardware, or serve low-latency requests. Those four groups have no path forward here. The absence of a metered price makes cost modelling impossible; the absence of benchmarks makes quality comparison impossible; the closed distribution makes independent testing impossible; and the mandatory thinking mode with a max default makes low latency a manual, per-call discipline rather than a setting.
The broader read: the specs that were published are genuinely competitive — 1M context and native video input in a coding model is a real capability, and the three-endpoint compatibility lowers integration cost. But the published specs are also the easy ones. The parts a buyer needs most were left blank, and the most user-hostile design decision — forcing reasoning always on while defaulting to the most expensive effort tier — is the one thing MiniMax stated clearly and unambiguously. That asymmetry is worth noticing.
If MiniMax publishes benchmarks, adds a metered rate, and relents on the thinking switch, this becomes a straightforward recommendation for budget-conscious long-context work. Until then, treat it as a capable-looking preview that has not yet earned the trust its spec sheet implies — and if you are on Token Plan, set effort explicitly on every call.
FAQ
When was MiniMax M3.1-Flash-Preview released?
It went live quietly on September 27, 2026 and entered public beta on September 28. There was no launch event, blog post, model card or pricing page. It arrived 120 days after MiniMax M3, which shipped June 1, 2026.
Is MiniMax M3.1-Flash-Preview the anonymous Space Bunny Alpha model?
Very likely, but unconfirmed. Space Bunny Alpha appeared free on OpenRouter and OpenCode on September 23, 2026 with a 1M-token context and text-image-video input. Multiple independent tokenizer probes — 24 test sets, 36 comparisons, 50 of 50 strings, 24 of 24 probes — all matched the MiniMax M3 family. A shared tokenizer identifies a family, not a version, so this is strong correlation rather than proof. MiniMax has never claimed Space Bunny.
Can you turn off thinking?
No. Sending thinking: {"type": "disabled"} or reasoning_effort: "none" returns HTTP 400 with error 2013. The only control is the five-level effort ladder — low, medium, high, xhigh, max — and it defaults to max when omitted.
How much does it cost?
There is no published per-token price. It is available only through the Token Plan subscription or MiniMax Code, with Token Plan tiers quoted at roughly $22, $55 and $132 per month. Cost can only be estimated as subscription divided by usage.
Why are there no benchmarks?
MiniMax published none, and third parties cannot produce their own because the model is reachable only through MiniMax Code. There is no entry on the Artificial Analysis index and zero recorded results on BenchLM.
What is the context window?
1,000,000 tokens, counting input and output together — the same ceiling as MiniMax M3.
Does it accept video input?
Yes. Video can be passed as a URL or base64 up to 50 MB. Images are capped at 10 MB and the request body at 64 MB.
Can it be self-hosted?
No. There are no open weights, which is a step back from M3. A restricted Hugging Face repository named MiniMax-M3.1-preview-private exists at roughly 250 GB, but it cannot be downloaded by outsiders.
How did the market react?
MiniMax stock fell roughly 9 to 10 percent on the day, to about HK$247.4 per share, reflecting continuing concern about revenue structure, losses and commercialization pressure.
Sources
- MiniMax M3.1 Flash Preview vs MiniMax M3: Which to Call — OrcaRouter
- MiniMax M3.1 Flash Preview Is Live on the Token Plan — OrcaRouter
- MiniMax released M3.1-Flash-Preview — DataNorth AI
- MiniMax M3.1-Flash-Preview: 1M Context, No Price Tag — Hokai
- MiniMax Quietly Slips M3.1-Flash-Preview Into Its Coding Tool — AlphaSignal
- MiniMax 首发只给自家 Code 和 Token Plan：1M 上下文、五档 effort，思考还关不掉 — 博客园
- 一次没有定价的发布：MiniMax M3.1-Flash-Preview 的匿名压测与静默上线 — 博客园
- 重挫10%！新模型和神秘"太空兔"，也没让MiniMax止跌 — 新浪财经
- 百万上下文、原生多模态！MiniMax新模型公测，疑似"太空兔"本尊 — 上观新闻
- 相隔120天，MiniMax终于上新 3.1 Flash Preview，结果股价暴跌9% — 东方财富
- MiniMax新模型硬核登场：思考过程强制开启，想"省事"没门 — 什么值得买
- MiniMax API Docs — Models Introduction
评论
发表评论