Skip to content
FrontierDot

MiniMax M3.1 Flash: From Oct 19 API Launch to Oct 23 Open Weights (428B / 23B Active)

Live timeline · last updated · 1 entry in the update log

Four days after MiniMax put M3.1 Flash on the API and said weights would come later, the same account posted that the weights were public: about 428B total parameters, 23B active per token.

MiniMax launched M3.1 Flash on its API on Oct 19, 2026 at $0.15 / $0.60 / $0.015 per million tokens and released the open weights on Oct 23 as a ~428B-total, 23B-active multimodal MoE under the MiniMax Community License.

Read on 2026-10-24 (UTC+8): MiniMax’s Oct 19 blog post, the Hugging Face model card for MiniMaxAI/MiniMax-M3.1-Flash and its repository API metadata, and official X posts by @MiniMax_AI on Oct 19 and Oct 23. X times are UTC as X reports them. Benchmark scores on the blog and card are MiniMax-reported. We did not run the model.

Timeline

  1. MiniMax publishes “MiniMax M3.1 Flash: Redefining Agent Performance and Cost,” dated 2026-10-19. The post says Space Bunny, which processed more than 200 trillion tokens while operating anonymously, was an earlier version of M3.1 Flash. It lists flat API prices of $0.15 input, $0.60 output and $0.015 cached input per million tokens, says the model is available through the MiniMax API, M Plan and MiniMax Code, and states: “Model weights will be released later.”

    MiniMax blog: MiniMax M3.1 Flash

  2. At 11:56 UTC the official @MiniMax_AI account posts about the M3.1 Flash rollout on X, matching the same-day blog announcement.

    MiniMax on X: M3.1 Flash API rollout (Oct 19)

  3. At 01:06 UTC @MiniMax_AI posts that MiniMax M3.1 Flash weights are now public, with 428B parameters total and 23B activated per token. A companion post follows the same morning. The Hugging Face repository MiniMaxAI/MiniMax-M3.1-Flash shows lastModified 2026-10-23T03:37:20.000Z on the model API the day we checked.

    MiniMax on X: M3.1 Flash weights public (Oct 23)

What shipped on October 19

The Oct 19 blog post is the product launch, not the weight dump. MiniMax says M3.1 Flash is a next-generation Flash model aimed at agentic tasks, coding, computer use and cybersecurity, and that online inference is nearly three times faster while usage costs drop by 70% versus its predecessor. Space Bunny is named in the same post as an earlier anonymous version that processed more than 200 trillion tokens. The post frames that anonymous run as product history rather than a separate model family: the same Flash line, previously unlabeled, now priced and named. That is useful context for anyone who watched Space Bunny traffic and wondered what it would become once MiniMax put a price sheet next to it.

Pricing is written as a flat table: $0.15 per million input tokens, $0.60 per million output tokens and $0.015 per million cached input tokens. The post says the model is live on the MiniMax API, M Plan and MiniMax Code, and ends the availability section with the sentence that matters for this timeline: model weights will be released later. A Pro model is mentioned as coming later as well; that line is a promise, not a release. For teams comparing Flash SKUs across labs, the useful check is whether the cache-hit rate on your workload is high enough for the $0.015 line to matter. MiniMax states the 70% usage-cost cut for agentic tasks against its own predecessor; it does not publish a third-party bill comparison on that page.

What the Hugging Face card says the model is

The model card describes MiniMax-M3.1-Flash as a native multimodal mixture-of-experts for coding, agentic workflows, computer use and cybersecurity. Inputs are text, image and video; output is text. The context window is up to 1,000,000 tokens. The card lists approximately 428B total parameters and 23B activated per token, MiniMax Sparse Attention (MSA) throughout the model, and a DSpark draft head for speculative decoding. The architecture class on the card is MiniMaxM3SparseForConditionalGeneration. Those design notes matter for anyone sizing a self-host: sparse attention and a draft head are MiniMax’s stated levers for long-context and decode speed, but they do not change the memory floor of keeping a 428B-class MoE resident. The active count tells you how much work each token does; the total count tells you how much VRAM or unified memory you still need to park.

The license is not Apache or MIT. The card sets `license: other` with `license_name: minimax-community` and points at a LICENSE file in the repository. Treat commercial use as gated by that MiniMax Community License until you have read the text. Front’s rule on custom licenses is the same as for Muse Spark Open: we do not paraphrase the clauses here. If you plan to fine-tune, redistribute, or serve the weights behind a paid product, the LICENSE file is the primary source, not the marketing line that the weights are public.

Hugging Face’s model API reports `createdAt` 2026-10-17T06:37:50.000Z for the repository and `lastModified` 2026-10-23T03:37:20.000Z. The public weight announcement on X is Oct 23, so this timeline dates the open release to that day. Why the repository object existed on Oct 17 is not explained on the pages we read. Empty or gated repos sometimes appear before a dump; we only treat the Oct 23 X posts and the Oct 23 lastModified stamp as the public release markers, and we leave the earlier createdAt in the unverified list rather than inventing a story for it.

Scores, as MiniMax reports them

Both the blog and the model card publish MiniMax’s own benchmark table. Highlights the card leads with include 76.2 on BankerToolBench, 73.7% on DeepSWE v1.1, 61.7 on OSWorld V2, 29.6 on Agents’ Last Exam and 82.2 on CyberGym. The blog also cites Automation-Bench and says output speeds can reach 200+ tokens per second in everyday use, with median (P50) speeds consistently reaching 170+ TPS. Those speed figures are operational claims from MiniMax, not measurements from an independent serving lab. Read them the same way you read the score table: useful as a vendor snapshot, not as a substitute for your own load test on the stack you actually run.

None of those figures are third-party verified on the pages we cite. They sit in the same category as every other provider score on Front: report what the company published, say whose number it is, and wait for an independent run before treating the ranking as settled. Side-by-side columns against DeepSeek V4.1 Flash-1017, Kimi K3, GLM-5.3 and closed models appear on MiniMax’s card; every competing number in that table is also presented by MiniMax, so the relative ordering is still a vendor chart until someone else reproduces the harness.

What we could not verify

  • Independent benchmark scores for M3.1 Flash. Every number on the blog and card is MiniMax-reported.
  • Why the Hugging Face repository shows createdAt 2026-10-17 when the public weight announcement is Oct 23. We date the open release from the X posts and the Oct 23 lastModified timestamp.
  • Hardware minimums to self-host the ~428B MoE. The card describes architecture and attention; it does not give a GPU floor we could check.
  • The full text of the MiniMax Community License beyond the card’s license_name and LICENSE link. Read the file before shipping a product on these weights.
  • What “Pro, which pushes the boundaries of intelligence, will also arrive later” means in practice. No Pro launch date was on the Oct 19 page when we read it.

Our take

The shape of this release is easy to follow: API and pricing first, weights four days later, with the same parameter story on X and on Hugging Face. The open question for builders is the license. MiniMax Community is not a permissive OSI license, so “weights are public” is not the same claim as Reflection Beam’s Apache 2.0 dump the same month. For API-only users the Oct 19 price sheet is the actionable page; for anyone pulling safetensors, the Oct 23 card and LICENSE file are. Keep this page as a live timeline; a Pro launch, a license clarification or a third-party evaluation each get a dated node.

Related reading on Front: the DeepSeek V4.1 Flash 1017 timeline covers another large Flash MoE that updated mid-October, the Reflection Beam timeline and the Meta Muse Spark Open timeline track the other major open-weight drops of the month, and every post is on the Front blog index.

Keep reading on Front

Every post, newest first, is on the Front blog index.

FAQ

When did MiniMax release M3.1 Flash weights?

October 23, 2026, per MiniMax’s official X posts that morning. The API and product launch was October 19, when MiniMax said weights would come later.

How many parameters does MiniMax M3.1 Flash have?

About 428B total parameters with 23B activated per token, according to MiniMax’s Oct 23 X post and the Hugging Face model card.

How much does the MiniMax M3.1 Flash API cost?

According to the Oct 19 blog post: $0.15 per million input tokens, $0.60 per million output tokens and $0.015 per million cached input tokens.

Is MiniMax M3.1 Flash open source?

The weights are public on Hugging Face under the MiniMax Community License (license_name minimax-community on the card). That is not Apache 2.0 or MIT; read the LICENSE file before commercial use.

Update log

  • — First published. Checked MiniMax’s blog, official X posts and the Hugging Face model card through Oct 24, 2026 (UTC+8).

Sources