On October 23, MiniMax announced the open release of M3.1 Flash model weights: roughly 428B total parameters, about 23B activated per token, a context window of up to 1M tokens, published on Hugging Face and ModelScope under a custom MiniMax Community License. Four days earlier, the model's anonymous predecessor Space Bunny had just been unmasked (official announcement of October 19: more than 200T tokens processed over 13 anonymous days); the open-source day added a fresh claim — over 100 trillion tokens handled within 72 hours of going live (official figures as relayed by media) — plus a roster of ten chip vendors signed up for adaptation. But one detail of this release is easy to miss: as of our verification date (October 24, 2026), GitHub hosts no code repository named after MiniMax-M3.1-Flash. What was opened is the weights, not the code. This piece walks through the weight repositories, the license terms, the empty GitHub result, and the actual readiness of the ecosystem list. The short version: the weights are genuinely downloadable, but the word "open" carries qualifiers that deserve a line-by-line read.
Scope and division of labor first: this article follows up on "MiniMax M3.1 Flash Unmasked: Space Bunny Ran 200T Tokens" (October 21, 2026), which covered the anonymous narrative, pricing, and the benchmark table. This piece handles only the open-source landing itself — where the weights live, how the license reads, why GitHub comes up empty, and which ecosystem commitments are real. Cross-model comparisons and the full local-deployment command walkthrough live in their own articles and are not repeated here. Source grading is marked throughout: the official model card was verified directly on ModelScope (both README and LICENSE pulled in full), the official announcement is quoted via media relay, and third-party figures are labeled as such.
1. Where the Weights Live: Two Repos, Two Org Names
Addresses first. ModelScope (verified reachable):
MiniMax/MiniMax-M3.1-Flash on ModelScope
Plain-text address: https://modelscope.cn/models/MiniMax/MiniMax-M3.1-Flash
Hugging Face (address endorsed by the official README download command; not directly reachable from our test network):
MiniMaxAI/MiniMax-M3.1-Flash on Hugging Face
Plain-text address: https://huggingface.co/MiniMaxAI/MiniMax-M3.1-Flash
Here comes the first trap: the organization names differ between platforms. On Hugging Face it is MiniMaxAI; on ModelScope it is MiniMax. If you type the HF-style MiniMaxAI/MiniMax-M3.1-Flash into ModelScope, you land on a shell page that returns HTTP 200 (a frontend fallback), while the API actually returns record not found. At verification time, the ModelScope page showed LicenseName as minimax-community, the registered architecture as MiniMaxM3SparseForConditionalGeneration, and 522 downloads — a plausible figure for a weight repo four days old.
The download command, quoted verbatim from the official README:
hf download MiniMaxAI/MiniMax-M3.1-Flash --local-dir MiniMax-M3.1-FlashOn size: the model card body does not spell out the weight format; an HF page snapshot (media-relayed) shows tags including an 8-bit precision marker. Third-party estimates put the repo at roughly 252GB across 48 safetensors shards, plus about 2.3GB for the DSpark draft head (community figures, not independently verified). Whether standalone FP8/BF16 variants or official quantizations shipped the same day: not announced; the README states only that published results correspond to the official weights and third-party quantizations may differ. For serving, the card recommends SGLang, vLLM (with an official recipes page), and TokenSpeed, with Transformers, KTransformers, unsloth, and ATOM integrations marked as in progress. Recommended inference settings are temperature of 1.0 and top_p of 0.95, with reasoning effort in five tiers from low to max; the README notes that usable context length and throughput depend on GPU memory and serving configuration. Multi-chip deployment instructions ship inside the HF repository docs directory as a FlagOS multi-chip support document. The full deployment walkthrough and memory budget are covered in our separate deployment SOP, linked at the end.
2. GitHub Comes Up Empty: Weights Only, No Code
This is the most honest thing to record about the release. Tested against the GitHub API on October 24, 2026: the URL github.com/MiniMax-AI/MiniMax-M3.1-Flash returns 404, and the organization page github.com/MiniMax-AI, while reachable, lists all 37 repositories with nothing related to M3.1 Flash — the newest model repos there remain MiniMax-M3 and MSA. In other words, as of verification time this open-source release consists solely of the weight repositories on HF and ModelScope, with no accompanying GitHub code repository.
Not surprising, but worth spelling out. "Open weights" and "open code" have long been two different things in this wave of model launches: a weight repo delivers the model files needed for inference, while inference framework code, training code, and data recipes are not part of the deliverable. M3.1 Flash's inference path depends entirely on third-party frameworks — vLLM and SGLang each wrote their own adaptation to load these weights, and those framework repos belong to the framework vendors, not MiniMax. The practical consequence: there is no official code to consult when inspecting the architecture definition, debugging load errors, or modifying the model; you work from the model card and reverse-engineer from framework sources. Comparable peers did it differently — DeepSeek V4.1 Flash 1017 and Reflection Beam both shipped with full code repositories, making MiniMax's weights-only format an outlier in this launch window. Whether a code repo will follow: the company has not said.
3. The License, Line by Line: Attribution and the $20 Million Line
The license is named MINIMAX COMMUNITY LICENSE (copyright MiniMax, 2026), pulled verbatim from the LICENSE file in the ModelScope repository. First, what it is not: not Apache 2.0, not MIT, but a custom community agreement. The key terms follow.
Non-commercial use is free, with rights to use, copy, modify, merge, publish, distribute, and sublicense — indistinguishable here from a traditional permissive license.
Commercial use carries two conditions. First, prominent attribution: a commercial product must prominently display "Built with MiniMax M3.1-Flash" on a related website, user interface, blog post, about page, or product documentation (the LICENSE's own wording). Second, the revenue threshold: any product or service from a company with more than 20 million US dollars in yearly revenue requires prior written authorization from MiniMax, requested via [email protected] with the subject line "M3.1-Flash licensing - authorization request"; commercial users under the threshold need only send a one-time notification email.
Two clauses are easy to skip. First, the definition of commercial use is broad: per the LICENSE, paid products or services count, commercial API resale (including cloud hosting) counts, and deploying the model for commercial purposes after any post-training or fine-tuning also counts — meaning fine-tuned descendants inherit the obligations; tuning does not launder them away. Second, the appendix prohibits military use, harm to minors, generating false or misleading information, and discriminatory or hateful content; the software is provided "AS IS" without warranty.
The contemporaneous contrast sharpens the picture: DeepSeek V4.1 Flash 1017 ships under MIT with zero addenda (covered in our unmasking piece and our license comparison). Against that backdrop, the MiniMax Community License is a textbook "usable with conditions" deal: small teams feel almost nothing (an attribution line plus one email), while large companies past the revenue line must run an authorization process. One accessibility detail worth noting: downloads are ungated on both HF and ModelScope — the conditions attach to usage, not to acquisition. Where this license sits on the openness spectrum is covered in our five-license comparison, linked at the end.
4. The Ecosystem List: Ten Chip Vendors, Three Tiers of Readiness
The official announcement (via media relay) names ten chip vendors: Huawei Ascend, Alibaba's T-Head, Kunlunxin, Moore Threads, Biren Technology, Hygon, MetaX, Iluvatar CoreX, NVIDIA, and AMD. Six inference and cloud platforms: Together AI, Fireworks, NeBIUS, Baseten, Modal, and DigitalOcean. Frameworks and communities: vLLM, SGLang, BAAI's FlagOS, Hugging Face, and ModelScope.
The list is long, but readiness splits into tiers and should not be compressed into a single "fully adapted" sentence. Per the media breakdown: Kunlunxin is the only chip vendor explicitly announcing a completed Day 0 adaptation; NVIDIA's NeMo team provides fine-tuning support, vLLM and SGLang handle inference, and Together AI is already serving the model; for Huawei Ascend, T-Head, AMD, Fireworks, and others, the official phrasing is "working with partners toward day-one access" — ongoing, not completed. Another set of media reports lumped everyone into "jointly advancing day-one access for the open-source launch," which reads as if everything has landed; both phrasings circulate, and this article adopts the more conservative tier. On the FlagOS side, BAAI announced Day 0 support for nine chips (FlagOS official figures), with the Ascend route running through its SGLang plugin. The one-line summary: this is a roster of announced partnerships; per-vendor verification will take more time.
5. Three Token Figures, Three Windows: 100T in 72 Hours, 200T in 13 Days, 56.7T on OpenRouter
The eye-catching number of the open-source day is "over 100 trillion tokens processed within 72 hours of going live." Two layers of sourcing apply: it is an official announcement figure relayed by outlets such as China Business Journal; and on the timeline, the model went live on October 19, so 72 hours lands on October 22 while the announcement ran on October 23 — roughly a day of ambiguity between the "72 hours" and the announcement date. The rigorous phrasing is "within 72 hours of launch (as of the announcement)."
This figure coexists with the two from our unmasking piece rather than competing with them: 200T+ cumulative over 13 anonymous days (official, October 19), 56.7 trillion tokens on OpenRouter during the anonymous window (relayed by ZhiDongxi, OpenRouter platform-side only), and 100T+ within 72 hours of public launch (official, October 23). Different windows, all valid; put "13 days for 200T" next to "3 days for 100T+" and the post-unmasking surge jumps out — the model moved more than half its anonymous-period volume in three signed days. Cite any of these only after checking which window it measures.
The cost claim needs the same alignment: the official wording is that comprehensive usage cost on agent-class tasks dropped 70 percent versus the previous generation, alongside "online inference speed up nearly 3x" and output above 200 tok/s (170+ at P50). For the record, a circulating "inference cost cut by half" claim has no traceable source and is not used here; "roughly half of M3's discounted price" is a third-party calculation that coexists with the official 70 percent figure. One more forward-looking line from the announcement (official, via media relay): a Pro version "pursuing the intelligence ceiling" is coming — no date, no specs, noted here for tracking.
6. Three Changes from M3: The Only Direct Official Statement
On how M3.1 Flash relates to its predecessor M3, the model card README is the only direct official explanation available, and its wording is: M3.1 Flash retains M3's core mixture-of-experts architecture with three main changes. First, sparse attention throughout: the first three full-attention layers are replaced with sparse attention, the mechanism being MiniMax Sparse Attention (MSA, paper arXiv:2606.13392). Second, quantization-aware training: the official announcement (via media relay) gives the details — routing experts at W4A4, with the sparse-attention indexer and main branch moving from Q8KV8 down to Q8KV4, which underpins the reductions in weight size and inference cost. Third, a speculative decoding swap: DSpark replaces the earlier EAGLE-like MTP approach, and the draft head ships with the weights, ready for third-party deployments to enable.
As for the 1M context window, the README's own sentence is that usable context length and throughput depend on GPU memory and the serving configuration — no minimum hardware requirements have been published, and there is no official consumer-grade single-GPU recipe. Readers who want the memory budget or the deployment commands should go straight to our deployment SOP, linked below.
FAQ
Q1: Where can I download the M3.1 Flash weights? A1: Two official repositories. ModelScope: https://modelscope.cn/models/MiniMax/MiniMax-M3.1-Flash (verified reachable); Hugging Face: https://huggingface.co/MiniMaxAI/MiniMax-M3.1-Flash (address backed by the official README download command). Note the ModelScope org is MiniMax, not the HF-style MiniMaxAI.
Q2: Is there a M3.1 Flash code repository on GitHub? A2: No. Tested via the GitHub API (October 24, 2026), the model-named repository returns 404, and none of the 37 repos under the official organization relate to M3.1 Flash. This release is weights-only; inference depends on third-party frameworks such as vLLM and SGLang.
Q3: What are the conditions for commercial use? A3: Under the MiniMax Community License, commercial use requires prominent "Built with MiniMax M3.1-Flash" attribution; commercial users with yearly revenue at or under 20 million US dollars send a one-time notification email, while those above the line need prior written authorization from MiniMax ([email protected]). Commercial deployment after fine-tuning still counts as commercial use.
Q4: Can I run it locally? What is the minimum VRAM? A4: No minimum hardware configuration has been published; the README states that usable context length and throughput depend on GPU memory and serving configuration. Official recipes start at multi-GPU scale (for example, eight H100-class cards), with no consumer single-GPU option. The command walkthrough and memory budget are in our deployment SOP.
Q5: How do 100 trillion, 200T, and 56.7T relate to each other? A5: Different measurement windows: 200T+ over 13 days is the official all-channel figure for the anonymous period; 56.7 trillion is the OpenRouter platform-side figure for the same window; 100T+ in 72 hours is the official figure after public launch. All three stand; check the window before citing.
Related reading
- MiniMax M3.1 Flash Unmasked: Space Bunny Ran 200T Tokens: the anonymous-period timeline, benchmark table, and pricing analysis — the prequel to this piece.
- Open Weight, But How Open? Five Model Licenses Compared: the openness spectrum from Apache and MIT to custom terms — the coordinate system for this article's license reading.
- MiniMax M3.1 Flash Local Deployment SOP: the full command-level walkthrough from download to multi-GPU serving, expanding this article's deployment notes.