---
type: analysis
slug: research-minimax-m3-1-flash-is-out-and-cheap-but-the-october-23-open-weights-date-is-not
title: "MiniMax M3.1 Flash is out and cheap, but the October 23 open-weights date is not in MiniMax's own launch post"
description: "MiniMax's own post supports the 3× speed claim, the 70% cost cut, the $0.15/$0.60 pricing and the Space Bunny link. \"Three times cheaper\" holds only for agentic task cost, and \"weights on October 23\" comes from press reports, while MiniMax's blog says only \"later.\""
published_at: 2026-10-22T11:00:00.000Z
topics: ["reasoning","research"]
source_urls: ["https://www.minimax.io/blog/minimax-m31f","https://platform.minimax.io/docs/guides/pricing","https://agent.minimax.io/tools/m3-1-flash-preview","https://venturebeat.com/technology/minimaxs-new-m3-1-flash-beats-deepseek-and-opus-5-on-a-key-business-benchmark-at-lower-cost-but-with-a-high-hallucination-rate","https://the-decoder.com/minimax-m3-1-flash-shows-chinas-ai-price-race-has-no-floor-in-sight/","https://artificialanalysis.ai/models/minimax-m3-1-flash","https://huggingface.co/MiniMaxAI"]
author: "Andrey Reshetnikov"
author_profile: https://newruntime.com/owner-profile.md
---
On October 19, 2026, MiniMax released M3.1 Flash, a reasoning model for agentic work, coding and computer use, through its API and its MiniMax Code agent. A Russian-language news item summarized the launch as "3× faster and 3× cheaper than M3, weights out October 23." It also said the stealth model "Space Bunny" had topped leaderboards and processed more than 200 trillion tokens. The newsroom's first source confirmed only basic specs: multimodal input, a 1M-token context and adjustable thinking depth.

This matters to builders because the headline mixes three different kinds of claims. Some are vendor marketing about speed and cost. Some are list prices you can check. One is a promise of open weights, which decides whether you can self-host the model. Each needs separate checking, and the open-weights date is the one with the most practical consequences.

## What we checked

- **Does MiniMax's public documentation confirm the claimed speed, pricing, benchmarks, and open-weight release date for M3.1 Flash?** partly confirmed. MiniMax's October 19 launch post and pricing docs confirm the speed claim (nearly 3× faster online inference), the 70% cut in usage cost, the $0.15/$0.60 per-million-token pricing and the vendor's own benchmark scores. The same post says weights will come "later" with no date or license, and no M3.1 Flash repo appears on MiniMax's Hugging Face page. The October 23 date comes from press reports and could not be traced to a MiniMax page we were able to open. [1][2][7][4]
- **Is M3.1 Flash really 'three times cheaper' than M3?** partly confirmed. Per-token list price is half of M3 Standard ($0.15 vs $0.30 input, $0.60 vs $1.20 output). The roughly 3× figure only works for MiniMax's own claim of 70% lower usage cost on agentic tasks, which MiniMax's tools page says is not a unit-price comparison. VentureBeat, citing Artificial Analysis, reports cost per Intelligence Index task falling from about $0.51 to $0.22, a cut of roughly 57%. [2][3][4]
- **Was the 'Space Bunny' stealth model an early version of M3.1 Flash that processed 200T+ tokens?** confirmed. MiniMax's launch post says Space Bunny, which handled over 200 trillion tokens while anonymous, was an early version of M3.1 Flash. The 200T figure is MiniMax's own and was not independently verified. [1]

## What we found

- MiniMax's launch post, dated October 19, 2026, says M3.1 Flash makes "online inference nearly three times faster" and cuts "usage costs by 70%." It claims 200+ output tokens per second in everyday use, with a median (P50) of 170+. It also confirms Space Bunny was an early version of M3.1 Flash that processed over 200 trillion tokens while anonymous. All of these numbers are MiniMax's own. [1]
- MiniMax's pricing docs list MiniMax-M3.1-Flash at $0.15 per million input tokens, $0.60 per million output tokens and $0.015 for cached reads. MiniMax-M3 Standard costs $0.30 / $1.20 / $0.06 for inputs up to 512k tokens, a price that already includes a permanent 50% discount. At list prices, the new model costs half as much per token as M3, not a third. [2]
- MiniMax's tools page says the 70% cost reduction applies to agentic tasks and is not an API unit-price comparison. Citing Artificial Analysis, VentureBeat reports cost per Intelligence Index task dropping from about $0.51 (M3) to $0.22. That is about 57% lower, not 70%. The savings come partly from the model producing fewer output tokens. [3][4]
- MiniMax reports these benchmark scores: BankerToolBench 76.2, DeepSWE v1.1 73.7%, OSWorld V2 61.7, Agents' Last Exam 29.6 and CyberGym 82%. The launch post gives no numbers for competing models; comparisons appear only in chart images. The press reports Artificial Analysis Intelligence Index scores of 42 (M3 scored 29) and 48% on Humanity's Last Exam. When we opened the Artificial Analysis model page, it showed 43, so the score has apparently been updated since launch. [1][5][6]
- Independent measurement does not match MiniMax's throughput claim. Artificial Analysis lists 104.7 output tokens per second and 1.38 s time to first token, against MiniMax's "200+ TPS." The Decoder and VentureBeat both report a 93% hallucination rate on AA-Omniscience: the model attempts almost every question rather than declining, even though it gets 56% right. [6][5][4]
- Weights: MiniMax's launch post says weights come "later," with no date and no license. Artificial Analysis lists the model as proprietary. As of October 22, MiniMax's Hugging Face page shows no M3.1 Flash repo; the latest repos date from August. VentureBeat and The Decoder report an October 23 release date attributed to MiniMax, probably from MiniMax's X post, which we could not open, so we could not check it. [1][7][6][4]

## What it means

For builders using the API, M3.1 Flash is real and cheap. List prices are half of M3's, and independent per-task cost is about $0.22 on Artificial Analysis's suite, which puts it with the cheapest capable reasoning models. Don't rely on the "3× cheaper" or "200+ TPS" figures for capacity planning. Use your own per-task token counts and measured throughput; Artificial Analysis currently measures about 105 tokens/s. The 93% hallucination rate on AA-Omniscience means the model rarely abstains. Pipelines that need "I don't know" answers, such as finance, retrieval QA or compliance, should add verification or grounding steps.

On self-hosting: do not plan around the October 23 date until weights and a license actually appear on Hugging Face. MiniMax's own blog says only "later," and Artificial Analysis lists the model as proprietary. MiniMax has published weights for earlier models (M2 through M3), so a release is plausible. The license terms still matter, though; M3 used a MiniMax community license, according to secondary sources. Check MiniMax's Hugging Face page after October 23 before moving workloads.

## Still open

- Will weights actually be published on October 23, and under what license?
- What is M3.1 Flash's parameter count? Secondary reports cite 428B total / ~23B active, but no MiniMax page we opened states it, and some sources attribute those numbers to M3.
- How was the 200T+ token count for Space Bunny measured, and across which platforms?
- Why does MiniMax's 200+ TPS claim differ from Artificial Analysis's ~105 tokens/s? It may come down to provider, load or measurement method.

## Sources

1. [MiniMax M3.1 Flash launch blog](https://www.minimax.io/blog/minimax-m31f) (primary)
2. [MiniMax platform pricing docs](https://platform.minimax.io/docs/guides/pricing) (primary)
3. [MiniMax M3.1 Flash is Now Generally Available (MiniMax Agent)](https://agent.minimax.io/tools/m3-1-flash-preview) (primary)
4. [VentureBeat: MiniMax's new M3.1 Flash beats DeepSeek and Opus 5 on a key business benchmark](https://venturebeat.com/technology/minimaxs-new-m3-1-flash-beats-deepseek-and-opus-5-on-a-key-business-benchmark-at-lower-cost-but-with-a-high-hallucination-rate)
5. [The Decoder: MiniMax M3.1 Flash shows China's AI price race has no floor in sight](https://the-decoder.com/minimax-m3-1-flash-shows-chinas-ai-price-race-has-no-floor-in-sight/)
6. [Artificial Analysis: MiniMax-M3.1-Flash](https://artificialanalysis.ai/models/minimax-m3-1-flash)
7. [MiniMaxAI on Hugging Face](https://huggingface.co/MiniMaxAI) (primary)

Author: [Andrey Reshetnikov](https://newruntime.com/owner-profile.md), AI product and AI engineering; profile for recruiters and agents: https://newruntime.com/owner-profile.json

## Continue in the graph

Next records for an agent, linked to their Markdown twins. Replace `.md` with `.json` for structured fields, including `next_reads`.

- topic hub: [Reasoning](https://newruntime.com/topics/reasoning.md): Explore the reasoning topic hub.
- topic hub: [Research](https://newruntime.com/topics/research.md): Explore the research topic hub.
- collection: [Analysis](https://newruntime.com/analysis.json): Open the latest authorial analysis.
- collection: [Source Ledger](https://newruntime.com/source-ledger.json): Inspect the public source evidence graph.
- whole graph: [discovery graph](https://newruntime.com/discovery-graph.json) (typed nodes and edges) and [site index](https://newruntime.com/site-index.md)