Open-weight model
An open-weight model has publicly released weights downloadable + runnable by anyone — Llama, Mistral, Qwen, DeepSeek, Flux Schnell / Dev, Stable Diffusion are 2026 open-weight families. Differs from open source (which would include training code + data).
Open-weight ≠ open source. Open weights means the trained parameters are publicly downloadable + usable under a license — but the training data, training code, and reproduction recipe may not be. Most 'open' LLMs in 2026 are open-weight, not fully open source: Llama, Mistral, Qwen, DeepSeek, Gemma, Phi for text; Flux Schnell / Dev, SDXL, Stable Diffusion 3 for image. Licenses vary: Apache 2.0 / MIT (broadest, commercial-friendly), custom (Llama license restricts > 700M MAU products), source-available (e.g., Flux Pro is API-only). Open-weight unlocks self-host, fine-tuning, embedded deployment, and security audit — and forces frontier labs to compete on quality + tooling rather than weights alone.
When to use open-weight model
- Self-host deployments.
- EU residency / data sovereignty.
- Fine-tuning for niche domains.
Common mistakes
- Treating open-weight as fully open source — license restrictions may still apply.
- Assuming open-weight quality matches frontier — closed models still lead on hardest tasks in 2026.
FAQ
What is open-weight model?
An open-weight model has publicly released weights downloadable + runnable by anyone — Llama, Mistral, Qwen, DeepSeek, Flux Schnell / Dev, Stable Diffusion are 2026 open-weight families. Differs from open source (which would include training code + data).
When should I use open-weight model?
Self-host deployments. EU residency / data sovereignty. Fine-tuning for niche domains.
What are the most common mistakes with open-weight model?
Treating open-weight as fully open source — license restrictions may still apply. Assuming open-weight quality matches frontier — closed models still lead on hardest tasks in 2026.
Related terms
- Self-hosted LLM — A self-hosted LLM runs entirely on infrastructure you control — your GPUs, your servers, your data residency — versus calling a cloud API.
- Model weights license — The model weights license defines what you can / can't do with downloaded model weights — Apache 2.0 (permissive), MIT (permissive), Llama Community License (restricted > 700M MAU), Gemma terms, FLUX-1-Dev non-commercial — production-critical to read before deploying.
Last updated: 2026-06-01. Raw markdown: https://promtable.com/glossary/open-weight-model.md.