DeepSeek V4 ships under MIT. NVIDIA's Nemotron Bolt ships under OpenMDW-1.1 with weights, training data, and recipe fully disclosed. StepFun's Step-5 ships under a license that says: cross twenty million dollars in revenue, and the weights stop being free. Four independent labs. Four independent legal teams. The same number. That convergence is not coincidence. It is the market discovering a price where open weights stop being marketing and start being leverage.
I have spent nine years watching protocols pretend that permissionless access is a technical property when it is almost always a legal one. Blockchain learned this the hard way โ GPL contagion in Bitcoin Core forks, the licensing fights around Ethereum clients, the quiet re-licensing of infrastructure codebases once the founders realized the moat was never the code. The AI labs are now walking the same path at ten times the velocity, and most of the industry is reading the announcement threads instead of the terms of service.
The license is the product. The weights are the distribution channel. Once you internalize that inversion, the October 15 release cluster stops looking like a model capability race and starts looking like a coordinated land grab conducted in legal language.
The three regimes, stripped of their marketing
Strip the branding and three distinct value-capture architectures visible in the filings.
Unrestricted (MIT / Apache 2.0). DeepSeek V4 and GLM-5.3 Flash. Zero license revenue. The entire return comes from ecosystem gravity, brand equity, and downstream services. This is the model that analysts reflexively call "generous." It is not generous. It is a bet that commoditization is coming anyway and that being the reference implementation is worth more than the license fee you would never successfully collect.
Revenue-gated. StepFun, Meta, Kimi, Mistral. The weights are free until a revenue threshold triggers a commercial negotiation. Here the details matter more than the headline. Mistral's Large 4 measures the threshold on a monthly revenue snapshot. StepFun measures on a rolling twelve-month window. Meta and Kimi sit between. Same twenty-million-dollar figure. Materially different clocks. A company with lumpy enterprise revenue can sit safely below a rolling annual line while blowing past a monthly one in a single strong quarter. That difference is not editorial sloppiness โ it is a designed conversion trigger.
Hardware-aligned. NVIDIA's Nemotron Bolt. The model is a loss leader for silicon. Every open-weight model that runs efficiently on NVIDIA hardware pulls demand for NVIDIA chips, so gating Bolt's adoption would be, in the cold language of platform economics, a landlord restricting the software that sells his own hardware.
Meta's position is the most legible of the four because it publishes its own tell. Muse Spark Open scores 39 on the Artificial Analysis Intelligence Index. The closed Wonder scores 56. Seventeen points. Roughly thirty percent relative capability gap. Meta calls this a structural moat. Read it more precisely: the open tier is deliberately positioned as the sub-flagship, calibrated to capture developers, seed tooling ecosystems, and generate the fine-tuning corpus that eventually feeds the paid tier. This is freemium rebuilt at the model layer, and anyone who has watched a SaaS funnel will recognize the geometry instantly.
Where the architecture actually binds
Here is the part the announcement coverage misses entirely. None of these four releases disclosed model architecture. No mixture-of-experts configuration. No attention variant. No context-length spec. No quantization roadmap beyond a single "BF16" annotation on Step-5. In a document set that runs to thousands of words of legal drafting, the technical specification is absent.
That absence is the finding. When an industry's flagship communication describes capability by license tier rather than by benchmark, the competitive axis has already moved. The labs are not competing on whether their models are better. They are competing on the terms under which you are permitted to find out.
And the terms are getting structurally aggressive. StepFun's clause 2c-i is the one to read twice. Device manufacturers โ phones, vehicles, glasses, wearables, robotics โ must negotiate a separate commercial license regardless of revenue. A two-person startup building an on-device assistant at zero revenue is in scope. This is not a revenue gate. It is a channel gate, and it is unique in the field. If it propagates, model licenses become hardware-selection locks. A deployer who builds on Step-5 cannot migrate to a more efficient silicon target without renegotiating, which means the hardware decision gets made by the legal team, not the performance team.
NVIDIA, by contrast, has no reason to gate anything. Its revenue arrives from the chip, not the weights. Full openness is not altruism โ it is asymmetric competition. Every CUDA-tuned open model reinforces the inference stack that competitors cannot easily replicate. Offer the same OpenMDW-1.1 license on an AMD or Ascend deployment and you inherit the license but not the optimization. The license is free. The moat is in the compiler.
What the trade actually costs
Weights at zero dollars is the least interesting number in the equation. The full cost of an open deployment is inference compute at production context lengths, compliance review across every jurisdiction you operate in, the legal exposure of a training-data provenance you cannot inspect, and the operational burden of running a serving stack nobody is paging at 3 AM but you. Open weights do not reduce total cost of ownership. They relocate it โ from a metered API invoice to a capital and headcount line item that most finance teams have never modeled.
I built a version of this spreadsheet during the Layer 2 benchmarking work, comparing rollup operating costs nobody wanted to publish against the marketing numbers everyone did. The pattern held: the free option is free at the layer you are looking at and expensive at the layer you are not. Two of the four open releases here carry no stated posture on derivative models under distillation. If "license follows derivative works" language hardens across releases โ and the drafting suggests it is heading that way โ the fine-tuning pipeline that most production teams depend on becomes a licensing audit surface.
The contrarian read: the twenty-million-dollar line is not protectionism. It is lead qualification. Below the threshold, self-serve adoption. Above it, you have stopped being a user and become an enterprise account, and the license restructuring is the opening move in a B2B sales cycle that never appears in the marketing collateral. The number is uniform because the market has converged on where small commercial use ends, not because four legal teams conspired. The absence of published post-threshold pricing is the tell. You cannot sell what you have already given away for free โ but you can start the conversation.
The audit nobody runs until it is too late
The chain is only as strong as its weakest node, and in an open-weight deployment the weakest node is almost never the model. It is the compliance posture of the team that chose the license.
Watch three signals. Whether Mistral's monthly measurement survives contact with enterprise buyers or gets quietly shifted to annual. Whether StepFun's OEM clause gets copied by any of the other three, which would convert a single vendor's aggressiveness into an industry default. And whether any deployer publicly documents a post-threshold negotiation, because the first published data point tells you what the gate is actually worth.
Code does not lie, but it often omits the truth โ and so do licenses. The weights are free. Read the next eleven pages before you believe it.