On October 15, 2026, three major open-weight models shipped within hours of each other. StepFun’s Step-5 BF16 landed after a same-day slip. Meta’s Muse Spark Open arrived with a gated download form. NVIDIA’s Nemotron Bolt shipped with weights, training data, and recipes fully open. The next day, NVIDIA’s RTX Spark hardware launched in the United States. Mistral Large 4, with open weights expected by month’s end, rounds out the quarter’s most concentrated open-weight window.
Four models, four licenses, one word: “open.” The term has decoupled from its legal meaning. What deployers actually received in this 24-hour window was not a single trend but a taxonomy of three distinct commercial strategies.
The first class is unrestricted. DeepSeek V4 ships under MIT. GLM-5.3 Flash ships under Apache 2.0. No revenue thresholds, no OEM gates, no MaaS definitions. Download, deploy, scale — the license does not care how much revenue you generate.
The second class is revenue-gated. This is where the industry is converging, and the numbers tell the story. StepFun’s Model License 1.0 imposes a $20M MaaS gate over any consecutive 12-month period, plus a hardware OEM gate no other lab has attempted. Meta’s Muse Model License applies the same $20M figure over any consecutive 12 months. Kimi K3 gates at $20M in annual MaaS revenue. Mistral’s Modified MIT License triggers at $20M in monthly revenue — a measurement that catches volatile businesses faster than a rolling aggregate. The number is the same. The clock is not.
The third class is hardware-aligned. NVIDIA’s Nemotron Bolt ships under OpenMDW-1.1 with zero deployment restrictions. No revenue gates. No OEM clauses. No MaaS definitions. The model is fully open because the company selling it makes money from the silicon underneath, not the model on top.
The $20M threshold is no longer coincidence. It is a convergence across four labs, each arriving at the same figure through independent licensing decisions. The measurement periods differ — monthly, annual, rolling 12-month — and those differences matter. A business with volatile revenue hits Mistral’s monthly snapshot faster than StepFun’s rolling aggregate. A steady-growth enterprise gets more runway from the 12-month window. But the trigger point is the same, and it is low enough that any significant inference-as-a-service business is in scope.
The practical effect is that “open weights” now means “downloadable,” not “deployable at scale.” The weights are free until your revenue makes them expensive.
NVIDIA’s incentive structure explains its permissive terms. The company earns revenue from GPU sales, not model access. Every open model that runs efficiently on NVIDIA hardware drives demand for NVIDIA silicon. A $20M revenue gate on Bolt would be the landlord restricting the adoption of software that sells its own hardware. The compute landlord gets vertical: the model, the silicon, the inference stack, and the license are all designed to work together as an ecosystem play.
StepFun and Mistral gate at $20M because their business model depends on capturing value from the model itself. NVIDIA’s business model depends on the model running on as many NVIDIA devices as possible. The permissive license is not generosity. It is infrastructure strategy.
Meta’s Muse Spark Open adds another dimension to the taxonomy. The model scores 39 on the Artificial Analysis Intelligence Index; the closed-source Wonder scores 56. That 17-point gap is not a limitation — it is a structural moat. Meta has explicitly bifurcated its portfolio: the frontier stays closed, the open track captures developers. The capability gap ensures that organizations requiring top-tier reasoning remain inside the proprietary tier.
For deployers, this means “open” models may intentionally lack the frontier performance needed for the most complex workloads. The gap is the product strategy.
StepFun’s hardware OEM restriction remains unique. Under Section 2c-i of the StepFun Model License 1.0, device manufacturers must negotiate a separate commercial license before embedding Step-5 in phones, vehicles, smart glasses, wearables, or robots. The restriction applies regardless of revenue. No other major lab — not Meta, not Mistral, not Kimi — restricts hardware embedding in this manner.
If this practice spreads, enterprise deployers could find themselves unable to move models to more efficient hardware, effectively tethering infrastructure strategy to the licensing terms of the model provider.
Licensing terms must now be audited with the same rigor as Mixture-of-Experts benchmarks or latency metrics. The questions are concrete: What revenue measurement period governs the license? Is there an OEM gate? Does the MaaS definition cover your specific service architecture? Does the license follow derivative works?
The Q4 open-weight window produced the industry’s clearest licensing taxonomy to date. The label “open” now spans three distinct commercial strategies, and the deployer cost of choosing wrong is not technical — it is legal and financial. The license is the product. The gate is the business model.
