What Open-weights model means in practice
An open-weights model is an LLM whose trained parameters are publicly available for download under a licence that permits commercial use. The major families in 2026 are Llama 3.3 (Meta, permissive licence with a 700M-MAU restriction that affects almost no enterprise), Qwen 2.5 (Alibaba, Apache 2.0), Mistral Large 2 (Mistral, commercial licence), DeepSeek V3 (MIT), Phi-3.5 (Microsoft) and Gemma 3 (Google, Apache 2.0). For sovereign deployments only open-weights models qualify — closed-API frontier models cannot be air-gapped because the weights never leave the vendor's infrastructure. Capability has converged: on the enterprise workloads we ship (Q&A, classification, extraction, summarisation, agentic orchestration), Llama 3.3 70B and Qwen 2.5 72B are within single-digit percentage points of GPT-4-class evals.
Related terms
Sovereign AI →
An architecture where customer data never leaves the network perimeter, model weights run on customer-controlled hardware, inference logs stay in the customer's SIEM, and the entire stack can operate air-gapped.
vLLM →
The high-throughput open-source inference server for LLMs — uses PagedAttention and continuous batching to serve open-weights models at production rates on a single GPU.
Large Language Model (LLM) →
A transformer-architecture neural network trained on very large text corpora to predict the next token in a sequence, producing fluent natural-language output across a wide range of tasks.
More in this category
All 62 terms, in plain language
Sovereign AI, RAG, agentic AI, IDP, MLOps and the regulations that shape enterprise AI.