Self-hosting an open-weight model settles where inference runs while leaving commercial rights governed by the company around the server. Engineering teams often treat downloadable weights as a one-time check completed the day a checkpoint lands on internal GPUs. When I read the licence files across eight open-weight launches announced in October 2026, only three used permissive terms without revenue or hosted-service gates. The other five conditioned rights on one or more of group revenue, affiliate services, distillation, or external access.
Those three permissive releases still match the habits engineers bring from open-source software. Reflection’s Beam shipped weights under Apache 2.0 on 16 October, DeepSeek V4.1-Flash-1017 shipped under MIT the next day, and NVIDIA’s Nemotron 3.5 Super VL shipped on 15 October under OpenMDW-1.1. Outside notice retention and defensive litigation clauses, all three stay indifferent to company revenue, corporate affiliates, or customer exposure.
Corporate structure alone can flip the legal baseline on the other releases before anyone writes application code. Under the Mistral Large 4 licence published on 29 October, if a company or any affiliate runs a model-as-a-service business, self-hosting the weights for any external commercial purpose requires a separate Mistral agreement with no revenue floor. Moonshot’s Kimi K3.1 Pro licence, published with its early-November weights after its October launch, applies that same group-wide gate above twenty million US dollars in combined twelve-month group revenue.
Customer exposure creates a second tripwire when an internal prototype becomes an external product feature. Kimi’s internal-use exemption ends once outputs or capabilities reach third parties, and self-hosted commercial products above one hundred million monthly active users or twenty million US dollars in monthly revenue must display Kimi K3.1-Pro on the user interface. MiniMax’s October M3.1-Flash licence requires prominent attribution and an email notice for any commercial use, plus prior written authorisation above twenty million US dollars in yearly product revenue.
Synthetic training data and small model size bring their own conditions. Meta’s 15 October Muse Spark Open licence exempts embedded product features from its twenty-million-dollar model-as-a-service gate, yet defines derivatives to include any student model distilled from its outputs, carrying its attribution and usage policy forward. Liquid’s compact LFM3 models, released on 27 October for local hardware, withhold commercial rights under LFM Open License v1.0 for corporate groups past its ten-million-dollar annual revenue threshold and terminate automatically on breach.
Software teams learned years ago to separate permissive open-source licences from network-triggered terms. Open-weight models add group revenue caps, affiliate checks, interface badges, and distillation reach-through to that split. Recording a licence once at download misses the product changes that cross those lines. Open-weight licensing now belongs in architecture review whenever an internal tool opens to customers, a product crosses a revenue tier, a student model trains on outputs, or an affiliate sells model access.