Image: MiniMax — M3.1 Flash (official product image: benchmark chart from the MiniMax announcement)
📌 KEY HIGHLIGHTS
- Weights released 23 October 2026: MiniMax followed its 19 October API launch with the M3.1 Flash model weights, fulfilling the "will be released later" promise in its launch post.
- Reported at 428B total / ~23B active parameters — widely cited but not confirmed by MiniMax itself, which has not published the exact parameter count.
- Day-one ecosystem: hosted inference via Together AI, Fireworks, Nebius, Baseten, Modal and DigitalOcean; framework support in vLLM, SGLang, Hugging Face and Baidu's FlagOS; chips from NVIDIA/AMD to Huawei Ascend and Moore Threads.
- "Space Bunny" revealed: the anonymous model that quietly processed 200 trillion+ tokens on leaderboards was an early M3.1 Flash build.
- Licence still undisclosed: the predecessor M3 shipped under a MiniMax community licence, not MIT/Apache — read the terms before self-hosting.
What happened?
On 19 October 2026 MiniMax rolled out M3.1 Flash, a reasoning model tuned for agentic work — the multi-step tasks where an AI uses tools, writes and tests code, and operates software on your behalf. The model went live on the MiniMax API, the M Plan subscription and the MiniMax Code desktop app at a flat $0.15 per million input tokens, $0.60 per million output tokens and $0.015 per million cached input tokens, with no surcharge for long context. MiniMax's announcement said the model weights "will be released later" — and independent coverage from The Decoder and SiliconANGLE pegged the date at 23 October.
That date has now landed. On 23 October, MiniMax released the M3.1 Flash weights, with reporting from ChainCatcher and RuntimeWire confirming the release and naming Together AI as the launch production-inference partner, so developers can call a hosted endpoint instead of managing their own GPUs. According to MiniMax's own figures, the model had already processed more than one trillion tokens from real users within 72 hours of launch. The release also closed a small industry mystery: the stealth "Space Bunny" model that topped routing-service leaderboards with 200 trillion+ tokens was, MiniMax confirmed, an early version of M3.1 Flash.
MiniMax M3.1 Flash open weights: specs, pricing and the competition
M3.1 Flash keeps M3's 1-million-token context window (up to 524,288 tokens of output) and accepts text, image and video input. Reasoning cannot be switched off; developers pick one of five effort levels — low, medium, high, xhigh or max. MiniMax claims inference is nearly three times faster than M3, with output speeds reaching 200+ tokens per second in everyday use (P50 above 170), and says agentic-task usage costs 70% less than its predecessor. Independent measurement backs part of the speed story: Artificial Analysis clocked 123 tokens per second on its hosted test, above its 116 average.
| Model | Size (reported) | Context | API price per 1M (in / out) | Weights / licence | Cost per completed task* |
|---|---|---|---|---|---|
| MiniMax M3.1 Flash | 428B total / ~23B active (reported, not vendor-confirmed) | 1M | $0.15 / $0.60 flat | Released 23 Oct; licence undisclosed | ~$0.22 |
| DeepSeek V4.1 Flash | 552B total (NVIDIA card) | 1M | $0.30 / $1.20 peak (half off-peak) | Open, MIT | ~$0.26 |
| MiniMax M3 (predecessor) | — | 1M | $0.30 / $1.20 (doubles above 512K) | Open, MiniMax community licence | ~$0.51 |
| Kimi K3 | 2.8T / 104B (Hugging Face) | 1,048,576 | $3.00 / $15.00 | Open, Kimi K3 licence | — |
*Per-task estimates from Artificial Analysis: M3.1 Flash finishes tasks more cheaply because it emits fewer output tokens, not because its per-token price is the lowest.
On benchmarks — all of them vendor-run, so read them as MiniMax's own scorecard — M3.1 Flash lands near frontier-level on agent-heavy tests: 62.8 on Automation-Bench, 76.2 on BankerToolBench, 73.7% on DeepSWE v1.1, 61.7 on OSWorld 2.0, 29.6 on Agents' Last Exam and 82.2% on CyberGym (vulnerability discovery and validation). Against DeepSeek's V4.1-Flash-1017 it splits the results: it leads on Automation-Bench (62.8 vs 60.5) and factual accuracy on AA-Omniscience (56% vs 43%), while DeepSeek leads on DeepSWE (77.0 vs 73.7) and the aggregate AI index (45 vs 42). Technical notes from the launch: full-model sparse attention, quantization-aware training (QAT) and power-aware scheduling are the optimisations MiniMax credits for the efficiency.
Image: Wikimedia Commons / Carl Lender, CC BY 2.0 — Illustrative image: data centre server racks of the kind used to host open-weight models
Why it matters for developers
This release is part of a visible shift in the AI market: the fight among budget-tier models is no longer just about benchmark scores — it is about cost per completed task. For anyone running AI agents at scale, where one job can mean dozens of model calls, MiniMax's pitch is that M3.1 Flash finishes work with fewer tokens at a flat price, undercutting DeepSeek on per-task cost even though both charge $0.15 per million input tokens at list price. With the weights now out, teams that could not send code or data to a China-based API for compliance reasons can self-host or use a third-party host such as Together AI — a workaround that matters to enterprises with data-residency rules. The same logic drives other open-weight stories this month, from our NVIDIA Nemotron 3.5 retrieval models explainer to the OpenAI Decisions API going generally available at $0.10 per million tokens.
There are two honest caveats. First, the licence: until MiniMax publishes the M3.1 Flash weight terms, the model is "weights released, licence pending" — and M3's precedent is a MiniMax community licence, not an MIT-style open grant. That is a weaker promise than DeepSeek's MIT-licensed V4.1 Flash, which teams can already download and modify freely (see our coverage of DeepSeek's $15 billion funding round for why the Chinese open-model race is accelerating). Second, the headline figures — 428B parameters, 1 trillion tokens in 72 hours, 70% cheaper than M3 — are MiniMax's own numbers; the benchmark table above is vendor-run, and independent verification has so far covered only inference speed.
Should you care?
✅ Care — and test it this week — if you run AI agents or code-generation workloads at scale: this is now one of the cheapest ways to get frontier-adjacent agent performance, at $0.22 a task by one independent estimate, with weights you can self-host and day-one support in vLLM and SGLang. ⏳ Wait if: your legal team needs the weight licence published first, you already serve DeepSeek's MIT-licensed V4.1 Flash happily, or you need absolute transparency on the 428B figure before committing infrastructure. The practical read: the API is the safe route today; self-hosting becomes the interesting route once the licence lands.
FAQs
Is MiniMax M3.1 Flash really open source?
Not exactly — "open weights" is the precise term. MiniMax released the model weights on 23 October 2026 so developers can download and run the model, but it has not published the licence terms, and its predecessor M3 shipped under a MiniMax community licence rather than an MIT or Apache-style open-source grant. Treat it as downloadable and inspectable, but not yet as MIT-open like DeepSeek's V4.1 Flash.
How many parameters does MiniMax M3.1 Flash have?
It is widely reported at 428 billion total parameters with about 23 billion active per token, but MiniMax itself has not published the exact count — treat the figure as reported, not vendor-confirmed. At that size, the weights matter most to inference providers and researchers with serious GPU capacity, not laptop users.
How much does the MiniMax M3.1 Flash API cost?
A flat $0.15 per million input tokens, $0.60 per million output tokens and $0.015 per million cached input tokens, with no long-context surcharge. Artificial Analysis estimates a completed task costs about $0.22, versus about $0.26 for DeepSeek V4.1 Flash — the saving comes from the model emitting fewer tokens per task, not from a lower sticker price.
What was Space Bunny, and why does it matter?
Space Bunny was an anonymous model that appeared on developer routing platforms in late September 2026 and quietly climbed leaderboards, processing 200 trillion+ tokens. MiniMax revealed it was an early build of M3.1 Flash — meaning the model's strong real-world usage data predates its official launch, which is why MiniMax can already quote trillion-token usage figures.
TAKEAWAY
MiniMax has turned M3.1 Flash from an API-only product into a downloadable model in four days — a fast, cheap agent workhorse with real developer momentum, held back from true open-source status only by a licence MiniMax still hasn't published.
Follow World Tech on TechByCountry for the next hourly update.
Sources: MiniMax — MiniMax M3.1 Flash: Redefining Agent Performance and Cost (primary, 19 Oct 2026) • ChainCatcher — MiniMax M3.1 Flash open source, 1T+ tokens in 72 hours (23 Oct 2026) • RuntimeWire — MiniMax opens M3.1 Flash weights, Together AI hosting (23 Oct 2026) • RITS / NYU Shanghai — MiniMax ships M3.1 Flash (20 Oct 2026) • MarkTechPost — MiniMax M3.1 Flash release details (19 Oct 2026). Checked 23 Oct 2026.
Comments
Post a Comment