Strategy

Open Weights Are Not Open Licenses: The Four-Case Spectrum Buyers Must Audit

By IdeaBosqueLast updated: November 4, 2026

Key takeaways

Kimi K3.1 Pro's full 2.8T-parameter weights shipped on Hugging Face on November 2, two days ahead of the promised date — under a custom Kimi K3.1-Pro License, and Artificial Analysis still listed the model as proprietary on November 4 (felloai's November 4 tracker). The labeling lag is itself the lesson: the open/closed boundary is now instrument-dependent.
Four named non-permissive license cases shipped in four months of 2026 — Qwen's stripped Qwen3.8-Max weights, Mistral's modified-MIT MaaS carve-out, Meta's attribution-bound Muse Model License, and now Moonshot's renamed K3 license. "Open weights" describes a delivery event, not a permission set.
Artificial Analysis projects the open-versus-proprietary Intelligence Index gap narrows from 12 points to 3 once the K3.1 Pro weights are counted (AA's projection, via felloai) — self-hosting can now buy near-frontier capability, but only under license conditions a buyer must read.
License conditions now appear at every scale: MiniMax's $20M yearly-revenue line, Mistral's MaaS carve-out, Meta's "Built with Muse Spark Open" display requirement — the permissive Apache 2.0 baseline covers only part of the chart.
K3.1 Pro costs $2.11 per Artificial Analysis Intelligence Index task against K3's $2.00 — a stronger model for the same money per task, not the same model for less (felloai); cheaper tokens are not cheaper tasks.

Moonshot published Kimi K3.1 Pro's full weights — 2.8 trillion total parameters, 104 billion active, 1-million-token context, ~1.56TB across 96 shards — on Hugging Face on November 2, two days ahead of its November 4 promise, under a custom Kimi K3.1-Pro License. The license is the Kimi K3 license word for word with the name changed, and Artificial Analysis still listed the model as proprietary on November 4, two days after the upload (felloai). Four months earlier, the same release-sheet read would have been a single open question — the release-sheet audit we published on October 25 carried "the license is unnamed" as the weights day's top open item, and asked a buyer to run a five-check audit when the weights landed (the weights-day method).

The answer arrived, and it is not "open." It is the fourth named non-permissive case in a pattern that now spans every major open-weights lab: Qwen stripped features out of Qwen3.8-Max's download (the stripped-release analysis), Mistral wrote a Model-as-a-Service carve-out into a modified MIT license (Mistral Large 4 open-weights announcement), Meta binds attribution into the Muse Model License (model card), and Moonshot renamed a restrictive license and shipped it with near-frontier benchmarks. This article maps the four cases on one spectrum, separates the license questions from the measurement questions a weights day settles, and gives the five-check audit to run before any self-hosting decision is priced.

The four cases, on one spectrum

The table below is the spectrum. Each release is a different mechanism for the same move — publishing weights while keeping a condition attached.

Release	License name	The condition	What a buyer must check
Qwen3.8-Max (Aug 2026)	Qwen3.8-Max License	Open weights are stripped: the download gets the base architecture and coding-agent benchmarks; the managed API keeps vision input, 1M context, thinking-mode control, and built-in tools	Compare the managed API against the capabilities the download actually carries — the model card and LICENSE file define the boundary
MiniMax M3.1 Flash (Oct 2026)	MiniMax Community License	A $20M yearly-revenue threshold plus attribution requirements gate commercial self-hosting (the weights-day audit)	Is your revenue — including affiliates — under the line? Price the API alternative at your workload's cache-hit rate
Mistral Large 4 (Oct 29, 2026)	Mistral Ecosystem Open-Weight license	Modified MIT: free to use, modify, fine-tune, and commercialize — with one condition: businesses operating a Model-as-a-Service offering must obtain a separate commercial license (Mistral's announcement)	Does your product make the model available to third parties? Internal use is exempt; the MaaS boundary is the license's single catch
Muse Spark Open (Oct 15, 2026)	Muse Model License	Limited commercial terms; "Built with Muse Spark Open" must be prominently displayed in the UI of each product or service using the model or its derivatives (LICENSE)	Attribution is a product-surface obligation, not a code comment — who ships your UI, and can they carry the display?
Kimi K3.1 Pro (Nov 2, 2026)	Kimi K3.1-Pro License	The Kimi K3 license renamed: internal use free; large Model-as-a-Service resellers need a separate agreement (felloai's license resolution)	Same check as K3. Watch same-license derivatives — the weights were pre-positioned under an "unrestricted" derivative listing before they landed

Three of the five rows are from the last three weeks. The pace is the finding: what was a per-lab quirk in August is a class in November.

The K3.1 Pro row resolves the release sheet's open item — as a restriction, not a permission. The October 25 release-sheet audit carried "the license is unnamed" as check #1 of the five-check weights-day audit, and priced it as the one item that could rewrite a self-hosting plan. The resolution is a renamed copy of the K3 license — a custom, MIT-derived document with commercial restrictions, including a mandatory separate agreement for large Model-as-a-Service resellers (AY Automate's license FAQ). Internal use is free (BERI's read). For a mid-market buyer whose agents answer internal workflows — a quote desk, an invoice-matching gate, a support graph — internal-use-free is the term that matters, and it holds. For a buyer embedding a frontier model into a customer-facing product, the separate-agreement clause is the whole story, and it applies at the licensee and any affiliate.

The Mistral row is the closest to permissive — and still not permissive. Mistral's own announcement states the carve-out plainly: the Mistral Ecosystem Open-Weight license is a modified MIT license, free to use, modify, fine-tune, and commercialize, "with a single condition: businesses that operate a Model-as-a-Service offering must obtain a separate commercial license from Mistral before using the model commercially" (announcement). For an enterprise running Le Chonk inside its own walls — or behind an embedded end-user feature that does not expose the model as a service — the condition may never bind. For a B2B software vendor whose product brokers model calls to customers, it binds immediately. The same license file, the same company, two opposite buying conclusions.

The Qwen row shows the second mechanism: condition by subtraction. Qwen3.8-Max's open weights shipped with the base architecture and the coding-agent benchmarks but without vision, beyond 262K context, with thinking locked (the stripped-release analysis) — the LICENSE file even covers both the stripped weights and the full managed service under one "Qwen3.8-Max License." Moonshot's K3.1 Pro is the contrast case in the same window: the November 2 download includes the full feature set — vision, 1M context, the works (felloai's specs table). Feature stripping and restricted terms are two ways to sell the same gap, and a purchase decision needs to name which one a release is running.

The spectrum in one image: where each release sits, the condition it carries, and the three numbers that decided most of it.

What the license file settles that the launch post does not

The spectrum has a permissive baseline, and it covers only part of the chart. Muse Glimmer — Meta's 30B dense agent model — ships under Apache 2.0 with no revenue cutoff (the bifurcation analysis); at that size class, permissive remains normal. Above the frontier, every recent open-weights release carries a named condition. The practical consequence is that "open weights" is no longer a binary a procurement team can inherit — it is a field to verify, per release, per condition type: revenue thresholds, MaaS carve-outs, attribution displays, feature stripping.

Two instrument facts sharpen the point. First, Artificial Analysis still listed K3.1 Pro as proprietary on November 4, two days after the upload — an indexing lag, not a capability fact (felloai). A buyer who queries an index for "is this model open" gets an instrument-dependent answer that trails the license file by days. And the same 55.1 Intelligence Index score appears labeled v4.3.2 on the K3.1-specific page and v5.0 on the November aggregate page — never cite the two labels together; cite one scale generation consistently, because the v5 scale and the v4 scale are not comparable numbers.

Second, the derivative risk is real before the weights even land: an "unrestricted" K3.1 Pro derivative was pre-positioned on Hugging Face ahead of the upload (the derivative listing). K3.1 Pro's restrictive terms do not travel into derivative checkpoints reliably — a distribution risk a self-hosting buyer inherits the moment a community fine-tune enters their stack rather than the vendor's checkpoint. Procurement that verifies license names at ingestion, not at announcement, catches this.

The second audit: what the weights day settles about the numbers

The license was the October audit's open item. The November weights day settles a second set — the dual-measurement rows that tell a buyer what the release sheet's claims become under independent harnesses. The release-sheet audit framed these as the checks that could not be answered sheet-first:

Terminal-Bench 4.0 ordering is harness-dependent. Artificial Analysis measured K3.1 Pro at 67% — the highest score it has measured on the bench — while Moonshot's own run printed 65.7% under the Claude Code harness, with Opus 5.5 at 66.4% on that harness (felloai's dual-measurement table). The 0.7-point gap against Opus 5.5 remains inside noise; the ordering against GPT-6 Astra (58.2%) holds in both columns.
The hallucination numbers agree in direction, differ in magnitude. Moonshot reports a fall from 23.4% to 8.3% on its internal stress bench; Artificial Analysis measured a smaller fall, from 53% to 44% (felloai). Both rows carry labels — the vendor's bench is a stress test, the AA row is a composite — and a quote that drops the label is a misquote.
Cost per task moved the wrong way while tokens got cheaper. K3.1 Pro uses 14% more output tokens per index task than K3, so Artificial Analysis' cost-per-task reads $2.11 against K3's $2.00 (felloai). The list price fell a third; the per-task price did not. This is the row a per-token spreadsheet misses.
The gap projection is now countable. With the weights out, K3.1 Pro is the highest-scoring open-weights model on the index and AA projects the open-versus-proprietary Intelligence Index gap narrows from 12 points to 3 (felloai). Three points is inside a benchmark's noise floor — which means the licensing question, not the capability question, is what now separates an open-weights deployment from a proprietary API in most mid-market buying decisions.

The five-check license audit, post-K3.1

The audit consolidates to five questions, each answerable from artifacts a buyer can read in an afternoon:

#	Check	The artifact	The failure it prevents
1	What is the license's exact name — and is it a standard one?	The LICENSE file in the weights repo	Pricing a self-hosting plan against Apache 2.0 assumptions on a custom license
2	Does any restriction key on revenue — and whose?	The license's threshold section, including affiliate language	The $20M-class line that converts "free commercial use" into "free until your next fiscal year"
3	Does your intended distribution fall inside a carve-out?	The MaaS/embedded-feature definitions, word by word	Discovering at product launch that embedding the model in a customer-facing product needs a separate license
4	Are the open weights feature-complete against the managed API?	The model card vs. the API docs, feature by feature	Building on a capability (vision, long context, tools) the download does not carry
5	What did independent harnesses reproduce?	Artificial Analysis / community harness rows beside the vendor's numbers	Quoting a 65.7% without the harness note, or a 8.3% hallucination rate without the stress-bench label

Run it in this order. Questions 1–3 are legal, and each one can end the self-hosting option before a single GPU is priced. Questions 4–5 are technical, and every one of the four cases above has at least one row a buyer would have missed by reading only the launch post.

A representative build

A mid-market wholesale distributor runs its quote desk on a managed proprietary model and wants the cost control of self-hosting its RFQ intake agent. The five-check audit runs in one afternoon: the target model's license is a named non-standard case (check 1 binds); the distributor's revenue sits far below any threshold, and the agent's answers surface only inside the company — internal use, no MaaS exposure (check 3 clears); the open weights carry the vision and long-context features the intake pipeline needs (check 4 clears); the independent harness rows reproduce the capability class within noise (check 5 clears). The self-hosted path is priced against the license it can actually run under — and the quote desk migrates its always-on retrieval workload onto the model the license permits, keeping the managed API for the two workloads with adversarial prompts where the vendor's stress-benched scores matter most.

Request a scoped build — we map the systems, the license and capability constraints, and the workflow in one week. You get a system inventory, a workflow map, and a fixed scope — whether or not you build with us.