Over the past seven days, the most consequential number in open-source AI was not a benchmark score. It was $20 million.
Kimi K3.1 Pro is live. The API is live, Kimi.ai is live, Kimi Work and Kimi Code are live, and there is already a "Highspeed" variant clocking two to three times the throughput of the default. The weights โ the actual artifact that the word "open" refers to โ are not. No license text, no repository, no terms. Moonshot AI's Hugging Face organization sits at 19 repositories with the newest update frozen at September 2. The model shipped. The permission did not.
Hype is the signal; silence is the warning. When a lab publishes a 3-trillion-parameter model and withholds the document that governs its use, the missing document is the most informative thing on the page.
Context: how "open" became a legal appendix
The modern open-weight era began not with a license but with a press release. Llama 2 in 2023 established the template: release the weights, attach an acceptable-use policy, add a commercial trigger, call it open. The community accepted the framing because the weights were genuinely useful and the trigger was set high enough that almost nobody hit it.
Kimi K2 followed a similar shape โ Modified MIT, with an obligation triggered at $20 million monthly revenue and 100 million monthly active users. The obligation itself was modest: display "Kimi K2" prominently in the product interface. An attribution requirement. Annoying, cheap, survivable.
K3.1 Pro is different in kind, if the reporting holds. The described obligation is not attribution โ it is a requirement to enter a separate negotiated agreement once the licensee crosses $20 million in group revenue, calculated on a trailing twelve-month basis, across all sources. That is not a license with a condition. That is a license with a door, and behind the door is a pricing conversation.
The technical claims attached to the model are substantial: 3T parameters, a 1M-token context window, and โ critically โ the disclosure that K3.1 Pro is a post-trained derivative of K3, not the base model. That last detail is doing more work than any benchmark in the release.
Core: the license is the tokenomics
In 2020, I spent a quarter mapping Curve's emissions schedule against the stability of its pools. The lesson that quarter taught me has never failed me since: the incentive schedule is the product, and everything else is packaging. A protocol's APY is not a yield โ it is a subsidy with a decay curve. Read the emissions, not the dashboard.
Open-weight licenses are the emissions schedule of the AI economy. Read them the same way.
First: the $20 million trigger is calculated on the wrong denominator. It is not model revenue. It is group consolidated revenue, trailing twelve months, all sources. A $25 million ARR software company that uses K3.1 Pro for a single internal summarization feature crosses the line. It then owes Moonshot a negotiation โ with no published ceiling on what that negotiation can demand: revenue share, minimum commitment, territory limits, or a clause governing whether your inference data flows back into the next training run. That is revenue-tiered licensing, structurally closer to an enterprise ELA than to an open-source grant.
Second: the "industry convergence" story does not survive contact with the license table. The framing that StepFun, Meta, Moonshot, and Mistral have converged on a $20 million MaaS threshold collapses on inspection. Meta's Llama community license triggers at 700 million monthly users โ roughly three orders of magnitude above the Moonshot line. Mistral's flagship tier has long been research/non-production. StepFun's Step series is Apache 2.0 by public record. And the two labs that actually define the permissive end of the spectrum โ DeepSeek with MIT, Alibaba with Qwen under Apache 2.0 โ are absent from the comparison entirely. The real structure is not convergence. It is bifurcation: absolute permissiveness at one pole, conditional access at the other, and Moonshot sitting mid-spectrum with a tightening bias.
Third: post-trained weights are not open weights. If only the K3.1 Pro checkpoint ships and the K3 base does not, downstream teams inherit an endpoint but not a starting point. They cannot swap the SFT dataset. They cannot redo alignment. They cannot audit what the safety filtering actually removed. What they receive is a self-hostable inference artifact โ valuable, but a different commodity than a researchable model. This is why a 3T checkpoint can be released at all without a strategic loss: what is being given away is the destination, not the road.
Fourth: the deployment arithmetic kills the democratization narrative before the license does. Three trillion parameters at FP8 is roughly 3TB of weights. A single 8รH200 node carries 1.1TB. Loading the model requires three to four nodes under tensor and pipeline parallelism, and a 1M-token context window adds KV cache in the tens to low hundreds of gigabytes on top. Realistically you are provisioning 64 cards or more. At current H200 rental rates, a single resident instance runs six figures per month.
That number matters for this industry. The decentralized-inference cohort โ Akash, io.net, Render, and their peers โ has spent two years pricing in the thesis that open-weight frontier models will create demand for distributed GPU supply. A 64-card floor excludes that cohort from hosting the head of the distribution. They will host the tail. The head stays with hyperscalers and sovereign compute platforms. The open-weight narrative and the decentralized-compute narrative are not the same trade, and the market has been pricing them as if they were.
Fifth: the safety accounting is unusually honest and unusually narrow. Moonshot disclosed that removing cyber-offense-related training data cost 9.5 points on its internal xdev-bench-hard โ 90.5% down to 81.0%, a 10.5% relative decline โ while Terminal-Bench 4.0 landed at 65.7% against a competitor's 66.4%. That 0.7-point gap is inside the noise band of any agentic benchmark and is vendor-reported besides. But the 9.5-point figure is one of the few public quantifications of an alignment tax in existence, and it deserves to be cited more than it will be. The caveat is that an internal benchmark tuned to the exact capability being removed will overstate the loss. It may also understate it: if the filtered corpus carried vulnerability mechanics, it carried defensive knowledge too, and security teams that wanted this model for code audit just lost capability they were not warned about.
Contrarian: the missing license is the product
The consensus reading is execution failure โ a promise made, a deliverable slipped. I think that reading is wrong, and expensive.
Consider the sequencing. API first, weights later. During the gap, every enterprise that wants K3.1 Pro's capability has exactly one procurement path: Moonshot's endpoint. No self-hosting, no third-party host, no negotiating leverage. That gap is the highest-margin window in the model's commercial life, and it is being harvested right now. Meanwhile Moonshot accumulates something no survey can buy โ real load data, real customer concentration, real willingness-to-pay curves โ which is precisely the dataset you need to set a revenue threshold and calibrate the terms behind the door.
I audited 40-plus token whitepapers in late 2017 for a Riyadh fund. The ones that failed did not fail on cryptography. They failed because the terms and the marketing described two different assets. I have watched the same gap open here, and the tell is identical: the promotional surface is fully specified, and the binding document is not.
There is a second, less comfortable point. The "we are three months behind the proprietary frontier, therefore releasing is safe" argument is self-authorizing. It measures risk in capability level and ignores diffusion velocity. A closed frontier model reaches users through a metered API with a refusal layer attached. An open checkpoint replicates at zero marginal cost, strips alignment with a fine-tune, and cannot be recalled. The risk structures are not comparable at equal capability, and the framing that they are has been repeated so often it now passes as governance.
And the compliance asymmetry should be familiar to anyone who has read a KYC policy. The license binds the party that reads licenses. A determined actor rents 64 GPUs and never signs anything.
Takeaway
Three signals will settle this, and none of them is a benchmark.
Whether the K3 base weights ship alongside the post-trained checkpoint โ that single fact separates "open model" from "open inference endpoint." Whether a third-party host appears within 30 days of release โ that is the timing and magnitude of the price shock. And whether any AI-agent token re-rates on the announcement โ if the market bids decentralized compute on news of a model it cannot host, the mispricing is confirmed.
Until the license text is public, treat every "open model" claim in this sector as an unpriced counterparty. The weights are the announcement. The license is the asset.