MiniMax M3.1 Release Date: M3.1 Flash Is Out, Open Weights Live, M3.1 Pro Pending

Updated 2026-10-26: M3.1 Flash has moved from a subscription-only preview to a general release with pay-as-you-go pricing and open weights. MiniMax has also confirmed that the "Space Bunny" stealth model on OpenRouter was an early M3.1 Flash build, and it has announced an M3.1 Pro with no date.

This page used to answer one question: when did MiniMax M3 come out? Since early October the more useful question has been whether M3.1 has shipped. It has, in stages. A gated MiniMax-M3.1-Flash-Preview showed up in MiniMax's API docs with no announcement. An unattributed model called Space Bunny ran free on OpenRouter for nearly two weeks (23 September to 5 October). Then on 19 October MiniMax launched M3.1 Flash properly, with a flat per-token price and weights on the way.

Below you'll find the release status of every M3-generation model, what changed from M3 to M3.1 Flash, what is actually known about M3.1 Pro (very little), and how the price change affects your budget. For full M3.1 Flash specs, API setup and serving recipes, see our MiniMax M3.1 Flash complete guide.

When was MiniMax M3.1 released?

MiniMax M3.1 Flash was released on 19 October 2026. MiniMax's model release notes carry an entry for that date: "MiniMax M3.1 Flash — An ultra-efficient open-weight model for everyday work." The launch announcement, as collected by Releasebot, says: "M3.1 Flash is now available through the MiniMax API, M Plan, and MiniMax Code." OpenRouter listed minimax/minimax-m3.1-flash the same day, and Unite.AI reports it was free there for the first 72 hours before switching to the standard price.

It came out in three steps, which is why release trackers disagreed for a few weeks:

1. Preview, undated. By early October, MiniMax-M3.1-Flash-Preview was documented in MiniMax's Anthropic-compatible API reference. It was available only through M Plan and MiniMax Code, with no pay-as-you-go price and no weights. We covered that state in the previous version of this page.
2. Stealth run, 23 September to 5 October. OpenRouter served an anonymous model listed as Space Bunny Alpha (stealth/space-bunny-alpha) for free. OpenRouter's models API showed it created on 23 September with a 1M-token context window and text, image and video input, as documented in our Space Bunny Alpha breakdown. Unite.AI reports it left OpenRouter on 5 October.
3. General release, 19 October. The model became M3.1 Flash, with a public rate card. MiniMax said weights would follow.

As of 26 October, MiniMax's models overview lists only two language models: MiniMax-M3.1-Flash and MiniMax-M3, both with 1M-token context. The Preview ID is gone from that page. If you hard-coded MiniMax-M3.1-Flash-Preview during the subscription period, switch to MiniMax-M3.1-Flash.

Are the M3.1 Flash open weights out?

Yes. The MiniMaxAI/MiniMax-M3.1-Flash repository is public and uses the same minimax-community licence as M3. That is MiniMax's own licence, not MIT or Apache 2.0, so read it before you deploy commercially. The sources give slightly different dates. Hugging Face's API reports the repository was created on 17 October 2026, two days before launch. According to Unite.AI, MiniMax's launch post said the weights would arrive on 23 October, and the announcement text just said "later". Repo creation and public visibility aren't always the same moment, so "mid-October" is the accurate summary.

Per the vendor's model card, M3.1 Flash is a natively multimodal Mixture-of-Experts model with about 428B total and 23B active parameters (the same footprint as M3), text, image and video input, and a 1M-token context. The main architectural changes are MiniMax Sparse Attention in every layer and a new DSpark speculative-decoding head. MiniMax recommends SGLang, vLLM or TokenSpeed for serving. (The Hugging Face Safetensors widget shows a lower count, likely because of packed U8 tensors; use the card's 428B for memory planning.) Full specs are in our guide's specs section.

The card doesn't publish hardware requirements. It only says usable context and throughput "depend on GPU memory and the serving configuration". Because the total parameter count matches M3, the SGLang M3 cookbook's validated single-node recipe is a reasonable starting point for capacity planning: 8 × NVIDIA B200 at --tp 8 with MXFP8 weights, or B300/GB300 at --tp 4. Our guide to running MiniMax M3 locally walks through the memory math.

Space Bunny was M3.1 Flash, now confirmed

Earlier in October we called Space Bunny a strong inference, not a fact. MiniMax settled it on 19 October: its launch post says Space Bunny, "which processed more than 200 trillion tokens while operating anonymously", was an early version of M3.1 Flash. OpenRouter's reveal, as reported by Unite.AI, adds platform figures:

• Space Bunny was OpenRouter's number-one model for 10 straight days.
• It processed 56.7 trillion tokens on OpenRouter, the most of any stealth model the platform has run.
• It left OpenRouter on 5 October and returned on 19 October as M3.1 Flash.
• Between the stealth build and the release build, OpenRouter credits user feedback with cutting time to first token from 1.95s to 1.22s, raising throughput from 70 to 100 tokens/sec, and halving the error rate from 1.4% to 0.7%.

The two token counts measure different things. 200T+ is MiniMax's figure and 56.7T is OpenRouter's, and neither source explains the gap. The likeliest explanation is that MiniMax's number includes traffic outside OpenRouter, but that is our inference, not something either company has stated. The full stealth-period evidence trail is in our space-bunny-alpha breakdown.

When is MiniMax M3.1 Pro coming?

There is no release date. MiniMax's 19 October announcement says only: "Pro, which pushes the boundaries of intelligence, will also arrive later." It gives no model ID, parameter count, context window, benchmark or price, and nothing called M3.1 Pro appears in MiniMax's models overview, rate card or Hugging Face org.

Two kinds of third-party claim are circulating, and both deserve caution:

• Placeholder dates. Some model catalogues show a specific M3.1 Pro release date. The ones we checked label it a placeholder, not an official timeline. Treat any exact M3.1 Pro date you see as a guess.
• The 2.7T figure is about "M3 Pro", not M3.1 Pro. In July 2026, AI Weekly relayed single-sourced reporting from The Information that MiniMax was preparing a 2.7-trillion-parameter "M3 Pro" for an open-source release "as early as Q3". MiniMax never confirmed it, and Q3 ended without that release. Whether the announced M3.1 Pro is the same project is unknown. This is an unofficial estimate, so don't size hardware or budgets around it.

If your roadmap depends on a larger MiniMax model, plan around M3.1 Flash and keep the model ID configurable.

What changed from M3 to M3.1 Flash?

The headline change is price. M3.1 Flash is half of M3's per-token rate, with no long-prompt surcharge. Rates below are from MiniMax's pay-as-you-go rate card as of October 2026:

M3's rates are still labelled "Permanent 50% off" against list prices twice as high. MiniMax-M2.7 is no longer on the pay-as-you-go rate card or the models overview, though OpenRouter still lists it. If you're still on M2.7, plan a migration now. MiniMax hasn't published a deprecation date that we could find.

Worked cost example

Take an agent workload of 10M input tokens and 2M output tokens per day, all on prompts under 512K and with no cache hits:

• M3.1 Flash: 10 × $0.15 = $1.50 input + 2 × $0.60 = $1.20 output = $2.70/day
• M3: 10 × $0.30 = $3.00 input + 2 × $1.20 = $2.40 output = $5.40/day

The gap widens on prompts above 512K, where M3's rate card lists a separate, doubled input and output tier (MiniMax's rate card doesn't spell out the billing rule, so check your invoices). Cached-input and long-context worked examples are in our guide's pricing comparison.

MiniMax's own marketing claims that agentic-task cost drops 70% and that M3.1 Flash costs one-third as much as M3. Those claims include token-efficiency gains from fewer and shorter reasoning turns, which you'll only see on your own traces. The per-token rate on its own is a 50% cut.

Benchmarks: vendor-reported vs independent

Vendor-reported, from the M3.1 Flash model card (MiniMax's own harnesses): M3.1 Flash leads its comparison set on Automation-Bench v1.0.6 (62.8) and ties Claude Opus 5 on BankerToolBench (76.2). It trails Opus 5 on Terminal Bench 4.0, OSWorld 2.0 and HLE, trails both DeepSeek V4.1 Flash (74.2) and Opus 5 (74.0) on DeepSWE v1.1 (73.7, and a newer DS V4.1 Flash-1017 build at 77.0), and trails GLM-5.3 on CyberGym (82.2 vs 84.5). The full table is in our guide's benchmarks section.

Independently measured, from Artificial Analysis on 26 October 2026, using Intelligence Index v4.3.2:

Speed and latency are live measurements and moved between our checks on 26 October, so they're shown as the range we observed; see Artificial Analysis for current values. The index and cost-per-task figures were stable.

That 14-point jump on the same index version is what makes M3.1 Flash a real upgrade and not just a price cut. Artificial Analysis gives M3 a 29 here, about what we reported on 5 October (29.22). Speed is another story. MiniMax says output speeds "can reach 200+ TPS", with median (P50) speeds "consistently reaching 170+ TPS". Artificial Analysis measures roughly 116–119 tokens/sec on MiniMax's own API. Plan capacity around the independent number.

When was MiniMax M3 released?

MiniMax announced M3 on 1 June 2026, which was the evening of Sunday 31 May US Eastern time. That's why some coverage dates it to 31 May. VentureBeat's launch piece is dated 1 June 2026, and MiniMax's release notes list M3 on 1 June.

The MiniMaxAI/MiniMax-M3 repository was created on 2 June 2026 at 07:49 UTC, and the quantised MiniMax-M3-MXFP8 repo followed a minute later. The MiniMax Sparse Attention paper, arXiv:2606.13392, followed on 11 June 2026. M3.1 Flash builds on that same paper. M3 is a ~428B-total / ~23B-active MoE with a 1M-token context window, according to its model card.

Here's where both generations sit in the MiniMax lineage. Dates are Hugging Face repository creation dates from the official MiniMaxAI org.

H3 isn't an M3 successor, even though the names look similar. It's a video model. If you're choosing a text, coding or agent model, the M line is the one to follow.

Should you move from M3 to M3.1 Flash?

• You're on M3 via the API. Test M3.1 Flash now. It costs half as much per token, has no 512K surcharge, and scores 14 points higher on Artificial Analysis's v4.3.2 index. MiniMax serves both models through the same platform and lists the same 1M window for each. Our MiniMax M3 developer guide covers the wiring, and it mostly carries over.
• You need thinking off for latency. M3.1 Flash can't do it. MiniMax's Anthropic-compatible API reference says it always thinks, returns a 400 error if you send thinking: {"type": "disabled"}, and defaults output_config.effort to max when omitted. Set output_config.effort (or reasoning_effort when self-hosting) to low explicitly, or stay on M3, which has thinking off by default. Benchmark latency on your own traffic before switching.
• You self-host M3. M3.1 Flash has the same parameter count, so your hardware plan carries over. Re-validate on SGLang or vLLM with the new DSpark speculative-decoding head before you cut over.
• You're comparing cheap open-weight agents. Put M3.1 Flash next to DeepSeek V4.1 Flash. They trade wins on MiniMax's own card. Our DeepSeek V4.1 Flash guide and fast, cheap LLM API comparison have the numbers, and our GLM-5.2 vs MiniMax M3 coding comparison covers the GLM side. For the wider open-weights field, see our open-source LLM landscape.
• You're waiting for M3.1 Pro. Don't block on it. Nothing about it is published.

M3.1 Flash is also available through M Plan subscriptions. As of October 2026, the regular monthly prices are $22/month (Go), $55/month (Explore) and $132/month (Build), as listed on MiniMax's monthly-offer page (the first-month discount on that page ended 14 October). For production traffic, the pay-as-you-go key and rate card above are the more predictable path.

FAQ

Is MiniMax M3.1 out?

Yes, the Flash tier is. MiniMax released M3.1 Flash on 19 October 2026 on its API, M Plan and MiniMax Code, priced at $0.15 input / $0.60 output per million tokens. Open weights are on Hugging Face as MiniMaxAI/MiniMax-M3.1-Flash. M3.1 Pro has been announced but not released.

What is the MiniMax M3.1 release date?

19 October 2026 for the general release of M3.1 Flash. A subscription-only MiniMax-M3.1-Flash-Preview was in MiniMax's API docs earlier, and an early build ran anonymously on OpenRouter as Space Bunny from 23 September to 5 October 2026.

When is MiniMax M3.1 Pro coming out?

MiniMax hasn't given a date. Its launch announcement says only that Pro "will also arrive later". There are no official specs, benchmarks or pricing. Exact dates in third-party catalogues are placeholders, and the widely quoted 2.7T-parameter figure comes from unconfirmed July reporting about a model called "M3 Pro".

Was Space Bunny made by MiniMax?

Yes. MiniMax confirmed on 19 October 2026 that Space Bunny was an early version of M3.1 Flash. OpenRouter says it processed 56.7 trillion tokens on its platform during the stealth run. MiniMax's own figure, more than 200 trillion tokens, likely covers a wider scope.

Can you self-host MiniMax M3.1 Flash?

Yes. The weights are public under the minimax-community licence. MiniMax recommends SGLang, vLLM or TokenSpeed. It's about 428B total / 23B active parameters, so plan for M3-class multi-GPU hardware. All the weights must be resident in memory, not just the active 23B.

When was MiniMax M3 released?

MiniMax M3 was announced on 1 June 2026, the evening of 31 May US Eastern time. Open weights were published to MiniMaxAI/MiniMax-M3 on 2 June 2026 at 07:49 UTC, and the Sparse Attention paper followed on 11 June 2026.

Is M3.1 Flash cheaper than M3?

Yes. As of October 2026, M3.1 Flash costs $0.15 / $0.60 per million input/output tokens at any prompt length. M3 costs $0.30 / $1.20 up to 512K and $0.60 / $2.40 above that. Cached reads are $0.015 on M3.1 Flash against $0.06 on M3 (under 512K).

Is there a MiniMax M3 Pro?

Not officially. The only official M3 variant is MiniMax-M3-MXFP8, a quantised build of the same model. A 2.7T-parameter "M3 Pro" was reported by The Information in July 2026 but never confirmed, and the announced M3.1 Pro has no published details.

Sources

• MiniMax: model release notes
• MiniMax: models overview
• MiniMax: pay-as-you-go pricing
• MiniMax: Anthropic-compatible API reference
• MiniMax: M Plan overview
• MiniMax: M Plan monthly offer (regular prices)
• Hugging Face: MiniMaxAI/MiniMax-M3.1-Flash model card
• Hugging Face: MiniMaxAI/MiniMax-M3 model card
• SGLang: MiniMax-M3 cookbook
• OpenRouter: models API
• Releasebot: MiniMax updates
• Unite.AI: MiniMax launches M3.1 Flash, revealing Space Bunny stealth model
• Artificial Analysis: MiniMax M3.1 Flash
• Artificial Analysis: MiniMax M3
• arXiv:2606.13392, MiniMax Sparse Attention
• VentureBeat: MiniMax M3 launch coverage
• AI Weekly: report on a 2.7T "M3 Pro" (unconfirmed)

If your team is moving agent workloads onto M3.1 Flash or self-hosting MiniMax weights, Codersera can help you hire vetted remote developers who have shipped LLM infrastructure in production.