Open weights as a go-to-market: the volume is won, the default is not

The next flagship from a fencing lab kept the fence, and the hosts came anyway. Moonshot opened Kimi K3.1 Pro’s weights in the first week of November, as it had promised at the model’s late-October launch (ITHome (opens in a new tab), 5 November): about 2.8 trillion parameters, 104 billion active, per the model card (opens in a new tab). Its licence (opens in a new tab) carries K3’s clause over word for word: a model-as-a-service business with more than $20 million in revenue over twelve months “must enter into a separate agreement with Moonshot AI”, unless it serves through Moonshot’s “official products or certified inference partners”. On 4 November Vercel wrote (opens in a new tab) that “until today, Moonshot AI was the only provider serving either model”; on 8 November its gateway lists 16 providers (opens in a new tab), and OpenRouter’s API returns 24 endpoints (opens in a new tab) from Together, Fireworks, Nebius, Alibaba and Moonshot itself among others.

Mistral’s Large 4, fenced with no revenue line at all, is the opposite case. Its weights (opens in a new tab) opened on 29 October under a “Mistral Ecosystem Open-Weight” licence whose text (opens in a new tab) says a licensee that “operates a Model as a Service business … must enter into a separate commercial license with Mistral before using the Model or its derivative works for any commercial purpose.” Ten days later Vercel lists one provider (opens in a new tab) for it, Mistral, at $0.70/$2.10 per million input/output tokens, and OpenRouter’s two endpoints (opens in a new tab) are both Mistral. Mistral’s own route for self-hosting is its platform: Studio (opens in a new tab) lets customers “deploy both the models and the platform in your own cloud, VPC or on-premises environment”, and BNP Paribas says it runs Studio on its own infrastructure (the post does not say which models); the licence exempts internal use. So two fences now behave differently: a revenue threshold lets the shelf fill, and a zero threshold has, so far, kept the serving with the lab. Demand muddies the comparison - Moonshot’s models took twice DeepSeek’s spend on Vercel’s gateway in September, and a lab’s pull on hosts is not the licence alone - so the counts show the fence holding, not the fence doing it.

Most labs fence the flagship; three do not

Read the licence files rather than the launch posts and most of the big releases since the summer carry a clause aimed at hosts. Kimi K3 (opens in a new tab) (July) wrote the template: the $20 million line, a definition of model-as-a-service as access with “meaningful control over the inputs, parameters, or training data”, and two exclusions, for products that embed the model in a feature and for “mere relaying of requests to models hosted by others”. GLM-5.3 (opens in a new tab) (August) uses the same definition and exclusions with a $10 billion bar and a security review, and Mistral’s Large 4 copies them too. An earlier version of this thread credited GLM-5.3 with them; Kimi K3 came first. Qwen3.8-Max (opens in a new tab) draws the line at $50 million and adds “AI Work Assistant” businesses. Meta’s Muse Spark Open (opens in a new tab) (15 October, 816 billion parameters, served from the first day by seven hosts (opens in a new tab) and the three big clouds) makes a host over $20 million “notify Meta and enter into a separate written agreement”. MiniMax M3 (opens in a new tab) needs prior written authorisation above $20 million in yearly revenue. Mistral had fenced before Large 4: its Medium 3.5 licence (opens in a new tab) (March) restricts any company over $20 million in monthly revenue, where Large 3 (opens in a new tab) shipped Apache 2.0 in December 2025.

Three labs ship flagships with no host clause. DeepSeek V4 Pro (opens in a new tab) (August, about 1.7 trillion parameters) and V4.1 Flash (opens in a new tab) are MIT, and so is Xiaomi’s MiMo-V2.6-Pro (opens in a new tab) (September, 1.02 trillion), which OpenRouter, citing Artificial Analysis, ranks as the best-scoring open model. Reflection AI’s Beam (opens in a new tab), 501 billion parameters with 23 billion active, is Apache 2.0; Reflection announced it (opens in a new tab) on 5 October as advancing “the Western open-weight frontier” and the weights followed in October. So the split is lab by lab and runs through the West as well as China: Meta and Mistral fence, Reflection does not. On 8 November Vercel’s gateway (opens in a new tab) lists four providers for it (Reflection, Baseten, Together and Nebius), while OpenRouter’s model list does not include it. Meta keeps both postures at once: its 30-billion-parameter Muse Glimmer (opens in a new tab) is Apache 2.0 while Spark Open is fenced.

What the fences collect is not public. No lab publishes a host agreement or its price. Vercel lists 18 providers for Kimi K3 (opens in a new tab) and 6 for the MIT MiMo (opens in a new tab); either most of Kimi’s hosts sit under $20 million, or they have deals, or nobody is enforcing, and no source says which.

The volume is won on two books; the money is on one

Vercel’s October index (opens in a new tab) puts open-weight models at 71.5% of September’s gateway tokens and nearly one dollar in four of spend, up from about one in twenty in June; its September report (opens in a new tab) had 56% of August’s tokens and 14% of spend. Each report says its open-weight classification is broader than the last, so compare within a report, not across them, and spend is estimated at list prices. It is Vercel’s own traffic, from a vendor that sells routing. No November index had appeared by 8 November.

OpenRouter gives a second count, from a different classifier (opens in a new tab) (a model is open if its catalogue entry links a Hugging Face repository): open-weight models took 27% of its top-50 tokens in the first week of October 2025 and 72% in the last week of September 2026, down from 77.5% in the week of 20 July. It leaves out stealth models - 38.7 trillion tokens in the latest week, unclassifiable - and it publishes no spend share. Its list prices put open models at about a quarter of closed ones per token ($0.28/$1.62 against $1.27/$6.36, token-weighted), and the best closed model scores about 11 points above the best open one on Artificial Analysis’s index. Two books from two routers agree on volume. The only split of money is still Vercel’s.

Inside that book, DeepSeek took 58% of September’s tokens and 6.6% of spend; Moonshot took 3.6% of tokens and 13.2% of spend, and the models priced at least four times the open-weight average, led by Kimi K3, accounted for about two-thirds of the rise in open-weight spend. The cheap MIT model is the volume; the fenced one is the revenue - on one gateway, at list price.

The devtools vendors have packaged the volume, cautiously. Vercel’s Token Pass (opens in a new tab) is $30 a month for $80 of credit on Kimi K3, GLM 5.3 and DeepSeek V4.1 Flash, all served by Wafer, drawn down at list price; on 8 November its docs still say Early Access with a waitlist, closed to Enterprise teams. On 6 November Vercel let the credit apply (opens in a new tab) to standard keys and OIDC tokens, which fall back to metered gateway credit when it runs out; Token Pass keys still stop. The flat plan can now feed the meter rather than stopping at it. OpenCode Go (opens in a new tab) at $10 mixes the same open models with small closed tiers. Alibaba Cloud’s Coding Plan (opens in a new tab), $50 a month for up to 90,000 requests, carries Kimi, GLM and MiniMax models beside Alibaba’s own Qwen Plus models, but the page does not say whether its Qwen models are open-weight, so Token Pass is still the one plan found that is priced on open weights alone.

The default is closed, and in one place it is closed by order

GitHub’s global model policy (opens in a new tab) for Copilot keeps open-weight models off by default regardless of what an admin sets, and the current docs (opens in a new tab) still name DeepSeek and Kimi K3 as out of scope on 8 November. SlashData and Mozilla’s survey (opens in a new tab) of 1,494 developers building AI features (May 2026) found 79% use open models against 71% for closed, but only 51% of open-model users reach production against 63% for closed; the closed-model rate climbs with company size and the open one does not. Self-reported, one wave, and Mozilla has named no date for a second.

The US government went further for one lab. A Department of Homeland Security order published on 20 October (opens in a new tab) excludes DeepSeek from federal civilian procurement and removes “all DeepSeek and DeepSeek-affiliate products, and third-party products and services developed using DeepSeek implementations” from civilian agency IT. The second half is the part that reaches a devtools vendor: an MIT licence lets anyone build on DeepSeek’s weights, and a federal buyer now cannot buy what they build. No order naming a second lab had been found by 8 November. A week after the order, Cohere said Alibaba’s Qwen3.8-Flash-Next tried to send a test string (opens in a new tab) to a server in Beijing during a security evaluation; critics said Cohere’s harness steered it, Cohere published its traces, and TechCrunch notes that Cohere sells its own models to governments. Nothing in it is a finding about Qwen; it is the kind of doubt the default is made of.

The weights still get into the enterprise, as one model among several or under another name. Databricks says the average customer on its platform now uses 7.1 distinct models, up from 4.2 in January, and that about 40% of them are open-weight (Databricks (opens in a new tab), 4 November; its own book, no customer named), and pitches a portfolio rather than a default. Microsoft post-trained Qwen3.5-9B (opens in a new tab) and sells it in Foundry as Microsoft-Decision-1. Ecosia moved its AI features from Mistral to Qwen, GLM and Kimi through the German host Melious, its chief executive saying the switch “roughly halved costs” (TechNode (opens in a new tab), citing POLITICO; second-hand). What the default keeps out is the open-weight model as the named first choice, not the weights.

The working reading

Open weights win the builder and the gateway’s volume. The labs releasing open flagships have split: most keep a claim on large hosts through a revenue threshold that leaves the shelf full, Mistral keeps the serving for itself and its enterprise platform, and DeepSeek, Xiaomi and Reflection leave it on the table. The enterprise takes the weights as one model in a portfolio or as a base it relabels, while its default stays a closed model - and for DeepSeek’s weights in US federal agencies, the default is now a rule. For a devtools vendor choosing a model to resell or build on, the licence file decides who it pays, the buyer’s policy decides who it can sell to, and the gateway price decides neither.

Each part is checkable: whether Large 4 gains hosts; a host publishing what a lab agreement costs, or a lab dropping the clause on its next flagship; Vercel’s next index on the spend share, and whether anyone else publishes one; a coding host flipping the default, or an order naming a second lab. If open-weight spend passes half on the gateway under an unchanged definition, or Copilot turns open weights on by default, the default half of this reading is wrong and the opening should say so.