Strategy

The $20 Million Line in the License: Weights Day Audits MiniMax M3.1 Flash

By IdeaBosqueLast updated: October 24, 2026

Key takeaways

MiniMax published M3.1 Flash's weights on October 23, 2026 — 48 MXFP8 safetensors shards in an Hugging Face repository stamped October 23, 03:37 UTC — four days after the October 19 API launch (Hugging Face model card).
The MiniMax Community License permits commercial use with attribution and a one-time notice below $20 million in yearly revenue — and requires prior written authorization from MiniMax above that line (license text).
The model card's benchmark table cuts both ways: M3.1 Flash leads DeepSeek-1017 on Automation-Bench v1.0.6 (62.8 to 60.5) but concedes Terminal-Bench 4.0 (38.1 to 42.3) and DeepSWE v1.1 (73.7 to 77.0) (model card benchmarks).
Artificial Analysis independently confirms the launch's cache economics — $0.15/$0.60 with a 90% cache discount, $0.22 per Intelligence Index task — and measures 43 on its index, against a 92.8%-confidently-wrong hallucination caveat (Artificial Analysis; Local AI Zone, aggregator-compiled).
Community quant conversions appeared within hours of the weights — five GGUF, MLX, and EXL builds created October 23 — and MiniMax ships first-party serving recipes for SGLang, vLLM, and TokenSpeed (Hugging Face; model card deployment notes).

Four days separate a model launch from its audit. On October 19, MiniMax launched M3.1 Flash as a price sheet and six benchmark numbers: $0.15 input, $0.60 output, $0.015 for cached input, and vendor-reported scores with no published harness. We carried the cache-rate argument the same day, and held two claims as unverified — the 428B parameter count existed only through a secondary write-up, and the weights were "released later" with the license undisclosed. On October 23 the weights appeared on Hugging Face and ModelScope, and what changed is not just availability: the release became an artifact a buyer can audit. This article reads that audit the way a procurement team should — license first, parameters second, format and ecosystem third, benchmarks with harness notes fourth — and adds the independent calibration column that a vendor table cannot print about itself. The conclusion for a mid-market buyer is specific: the audit writes a $20 million revenue line into the license terms, moves one benchmark lead into an open question, and leaves the cache arithmetic exactly where it stood.

Weights day, read as an audit: four artifacts

The license is the first artifact, and it carries the number most coverage skipped. The MiniMax Community License grants free use for non-commercial purposes. Commercial use is permitted, with two conditions in sequence: display "Built with MiniMax M3.1-Flash" on a related website, interface, or product documentation, and either send a one-time notice to api@minimax.io — or, if the products and services using the software generate more than $20 million (US) in yearly revenue, obtain separate prior written authorization from MiniMax before commercial deployment. The text also lists prohibited uses: military purposes, harm to minors, and disinformation among them. Read as a procurement document, the license does what launch posts never do — it segments buyers by revenue and states the paperwork each segment owes. A company at $150 million in annual revenue is not negotiating from the same position as one at $19 million, and the terms say so explicitly.

Parameters moved from secondary to primary. The model card states "approximately 428B total; 23B activated per token" — the count our launch-day article could only attribute to MarkTechPost's reporting while the primary stayed silent. The card also states what the launch post never did about architecture: sparse attention throughout, with the first three full-attention layers replaced; quantization-aware training; and a DSpark speculative-decoding draft head included in the release. Recommended sampling settings and the reasoning-effort ladder (low through max, thinking always on) are printed on the card.

The format and the ecosystem are the third artifact. The repository ships the weights as 48 safetensors shards with quant_method: mxfp8 in the config — the 8-bit format quantization-aware training was designed for. The Hugging Face API metadata shows the repository's weights stamped October 23 at 03:37 UTC, with 5,789 downloads and 303 likes by our verification pass on October 24. Community conversions followed within hours: unsloth's GGUF build and a second unsloth release, a community GGUF, two mlx-community 4-bit builds for Apple silicon, and an EXL3 quant all carry October 23 creation dates — the same pattern the predecessor M3's open-weights drop produced. MiniMax also published first-party serving recipes: an SGLang cookbook, a vLLM recipe, and TokenSpeed support, plus FlagOS multi-chip deployment notes. Conversion pace is a community-commitment signal — five independent builds inside a day is the ecosystem behaving as if the weights matter.

The benchmark table is the fourth artifact, and it cuts both ways. The card's table now carries harness notes the launch post lacked — the gap Local AI Zone's deep dive called "a launch post, not a model card." The columns the card adds are real wins: Automation-Bench v1.0.6 at 62.8 leads DeepSeek-1017's 60.5 and Claude Opus 5's 50.3 (600 publicly released tasks), LOCA-Bench 256K at 60.3 leads that column, and BankerToolBench at 76.2 ties Opus 5's 76.2 as published by the card. The columns that concede are just as real: Terminal-Bench 4.0 prints 38.1 against DeepSeek-1017's 42.3 and Opus 5's 51.8; DeepSWE v1.1 prints 73.7 against 1017's 77.0; Agents' Last Exam prints 29.6 under Opus 5's 32.2; and CyberGym's 82.2 trails GLM-5.3's 84.5 in the same table. The harness notes themselves deserve procurement attention: BankerToolBench is judged by a MiniMax model ("MiniMax-M2.7 judge"), a self-judging disclosure the launch post never made, while Terminal Bench runs under Codex 0.149.1 and DeepSWE under mini-SWE-agent 2.4.5 at 512K context. The card closes with its own honesty marker: "Results are for the official weights; third-party quantizations may differ."

The column the vendor cannot print: independent measurement

Artificial Analysis' model page adds the pricing verification a buyer cannot do from the vendor sheet alone: $0.15/$0.60 confirmed, with the cache discount independently read at 90% — the published $0.015 cached-input rate holding under third-party measurement four days after launch. The page also prices the agent's work: $0.22 per Intelligence Index task, where Local AI Zone's October dispatch reports the predecessor M3 at $0.51 per task (aggregator-compiled). The index reading is 43 — fifth of 173 models in its class — with measurable side effects a cost model should carry: the Intelligence Index run emitted 180M output tokens against a 100M median (verbosity is a bill, as the Sonnet 5.5 effort-economics analysis showed), and output speed of 95.6 tokens per second sits below the 111.3 median.

The calibration caveat is the sharpest number in the file. AA-Omniscience — 6,000 questions across 42 topics in six domains, where the methodology defines hallucination rate as incorrect answers divided by all non-correct responses and rewards abstention — scores M3.1 Flash at an index of 16, with 56% raw accuracy and 92.8% of its wrong answers confidently wrong rather than hedged or refused (compiled by Local AI Zone from Artificial Analysis measurements reported through VentureBeat and aimlapi — an aggregator-compiled row). For scale, the leaders on that index print 46, 44, and 43 — Claude Opus 5.5, GPT-6 Astra, and Claude Fable 5.1 respectively. The operational read is the one the dispatch draws: excellent for verifiable, structured work — code execution, tool calls, agentic loops with checkable outcomes — and the model most likely to invent an answer among everything tracked this month on open-ended factual retrieval without a verification layer. One honest footnote: the AA page still labels the model proprietary, a status the October 23 weights superseded.

What weights day changes for procurement

Read the license before the benchmark table. The M3.1 Flash license state, in two rows: below $20 million in yearly commercial revenue, the cost of self-hosting is an attribution string and a one-time notice; above it, the cost is a prior written authorization request to api@minimax.io — a consent gate with an unpredictable timeline, and a questions-for-counsel item your contract review should flag before the benchmark table wins anyone over. Neither condition applies to calling MiniMax's own API; the license governs the weights, not the vendor's hosted service. The distinction matters because the hosted path is the one where the published cache row applies anyway.

Hostability is not the same as hosting. The 428B-backbone hardware math is unforgiving at any quantization: Local AI Zone's planning table puts MXFP8 at roughly 428 GB — an 8× H200 or 4× B200-class server — Q4-class GGUF at 220–240 GB on a 512 GB-RAM workstation, and Q2-class builds flagged as not recommended for agentic work (Local AI Zone deep dive). For most mid-market teams without GPU infrastructure, the API remains the rational serving default; what the weights buy is not a hosting plan but pinned checkpoints, auditability of what you serve, and an exit path — the same substitute-risk discipline as the model-swap article prescribes. Hold the pinned checkpoint in storage even if you serve nothing on it.

The audit generalizes, and the calendar will give you chances to run it again. Five checks cover any weights day: license terms and revenue thresholds; the parameter statement's source (card or secondary); the quantization format and serving recipes; benchmark deltas with harness notes; third-party measurement including calibration, not just capability. The open-weight field is about to hand out more test cases — Kimi K3.1 Pro's full weights are due November 4 and Mistral Large 4's October 27, per the October dispatch (aggregator-compiled) — and the five checks run on each of them the same way. The flash-tier table our launch article printed carries a "weights" column; for MiniMax, that column finally has a row a buyer can verify, and the two claims we held on October 19 — parameters and weights — have moved into the confirmed set.

The one-minute diagram below puts the audit on one page:

The two panels are the audit in one page: the left side is what a buyer could check on October 19, the right side is what became checkable on October 23, and the violet card is the independent column no vendor table can print about itself.

A representative build

A mid-market industrial distributor, about $150 million in annual revenue, runs a quote-desk agent fleet that answers roughly 800 intake and fulfillment tasks a day, routed mostly to flash-tier models with the team's own MCP modules fronting the ERP and catalog layer. On weights day the platform team runs the five-check audit instead of a model swap: the license review flags the $20 million line (their revenue sits above it, so commercial self-hosting would need prior written authorization — the API path changes nothing); the parameter and format checks confirm the pinned checkpoints are retrievable at MXFP8 with first-party serving recipes if the exit path is ever exercised; the benchmark deltas with harness notes move Terminal-Bench-sensitive coding tasks onto DeepSeek-1017 while BankerToolBench-class document work stays; and the calibration row routes open-ended retrieval questions through a frontier model with a verification layer rather than trusting the flash tier's confidently-wrong 92.8%. The team logs the served model on every response, keeps the pinned checkpoint in storage, and re-runs the audit on the next weights drop.

Request a scoped build. One-week discovery. You get a system inventory, workflow map, and fixed scope — whether or not you build with us.