NVIDIA has crossed a threshold that changes the licensing conversation. The company that sells the silicon, the interconnect, and the inference stack now ships its own model, optimized for its own hardware, under terms that contain zero deployment restrictions.
On October 15, NVIDIA released the Nemotron 3.5 Bolt 74B-A9B on HuggingFace under the OpenMDW-1.1 license. The release coincides with the RTX Spark hardware line launching in the United States on October 16, with laptops from Microsoft and HP starting at $2,599 and a developer box at $5,999. The timing is deliberate; the model is a precision tool engineered for that specific hardware.
The license is the story
The most striking aspect of this release is not the architecture but the terms. OpenMDW-1.1 is NVIDIA’s own license — permissive, but not OSI-approved. It is neither Apache nor MIT. Critically, it is empty of the deployment restrictions that now define the open-weights landscape.
Deployment restriction convergence has become the new normal. StepFun’s Model License 1.0 gates OEM hardware embedding and Model-as-a-Service operations at $20 million in combined revenue over any consecutive 12-month period. Mistral’s modified MIT license triggers a commercial license requirement at $20 million in monthly revenue. Kimi K3 sets a $20 million annual MaaS gate.
NVIDIA’s license contains none of these. No revenue gates. No OEM restrictions. No MaaS definitions. The company with the most to lose from commoditized inference chose the most permissive terms on the market.
The model and the machine
Bolt 74B-A9B uses a mixture-of-experts architecture, specifically a Mamba-2-Transformer hybrid LatentMoE, with 74 billion total parameters and 9 billion active per forward pass. Native 2-bit IQ2 quantization compresses the model to fit within 24GB of memory, enabling it to run locally alongside other applications on RTX Spark hardware. The context window extends to 256,000 tokens.
The model was compressed from Nemotron 3.5 Super VL using Star Elastic architecture search, trained on 80 billion compression tokens on top of a parent corpus exceeding 20 trillion tokens. Benchmarks place it at 86.43 on GPQA Diamond, 66.75 on BrowseComp, 50.86 on WebArena (vision), and 87.82 on PinchBench — the top open model in its class on the latter. The NVIDIA blog announcement confirms training data and recipes are also fully open on HuggingFace.
The structural play
NVIDIA’s incentive structure is unique. 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 $20 million revenue gate on Bolt would be the landlord restricting the adoption of software that sells its own hardware.
StepFun and Mistral gate at $20 million 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.
The forward signal
The licensing taxonomy now has three classes. Restricted open-weight: StepFun, Mistral, and Kimi K3, gating commercial deployment at $20 million thresholds. Genuinely open: DeepSeek V4 and GLM-5.3 Flash, shipping under standard MIT or Apache 2.0 with no restrictions. And hardware-vendor open: NVIDIA, with a custom permissive license designed to maximize model adoption on its own hardware stack.
For deployers evaluating the Q4 open-weight landscape, the question is no longer just what the license says. It is who wrote it and why.