Open-Weight Models
Models whose weights you can download and run yourself — often called open source, usually not quite.
Reviewed July 11, 2026Stable
When not to use it
- When a hosted API would do. You're taking on infrastructure, evaluation, and updates to save money you may not be spending yet.
- Without reading the licence. Several popular "open" models carry user thresholds or use restrictions that rule out common cases.
- Assuming open means auditable. Without training data you can observe behaviour and nothing else.
Reach for something else instead
- Hosted APIs — better models, no operations, and the arithmetic favours them longer than people expect.
- Private cloud deployment of a hosted model, when the concern is data residency rather than cost.
- A smaller task-specific model when the real need was one narrow job, not general capability.
Read more on the blog
- Open weights vs open source AI: what's actually releasedAlmost every model marketed as open source AI is not open source. You get the finished weights under a licence, not the training data or the code that produced them. The difference decides what you can actually do, and regulators have now attached legal exemptions to the word, which makes the definition worth money.
- Run an LLM locally: how much VRAM do you need?Whether a model runs on your machine is not a mystery. It is one line of arithmetic: parameters times bytes per parameter, plus the context cache almost everyone forgets. This guide gives you the formula, the numbers for every common model size, and the two traps that cause most out-of-memory errors.
- How to read a model releaseEvery few weeks a lab announces a new frontier model and every headline says the same thing. Here's how to work out what actually changed, what the benchmark numbers mean, and which parts of the announcement are marketing.
Further reading
- Touvron et al. (2023), Llama 2: Open Foundation and Fine-Tuned Chat Models — the release that made this mainstream, licence and all.
- Solaiman (2023), The Gradient of Generative AI Release — the spectrum from closed to open, framed clearly.
- Widder, Whittaker & West (2023), Open (For Business): Big Tech, Concentrated Power, and the Political Economy of Open AI — the argument that "open" is doing work here it wasn't designed for.
Primary sources, listed so you can check the claims on this page rather than take them on trust.
Where people go wrong
- Calling them open source. Weights without data or training code is a compiled binary, not source.
- Underestimating the operational cost. GPUs, scaling, uptime, and updates are a team's work, not a weekend's.
- Skipping evaluation because the API provider used to do it. That job is yours now.
At a glance
FieldTools & Ecosystem
Core ideadownload the weights, run them yourself
Notopen source, in most cases
Real reasonsdata residency, volume, offline
Real costoperations and evaluation
DifficultyIntermediate
Often compared with
Open-weight vs. hosted API — your hardware and your quality bar vs. someone else's, at someone else's price.
Where this sits
28 concepts come first. Understanding it opens up 5 more.
11Levelsteps in
28Needs firstconcepts
5Opens up2% of 310
3Areascrosses fields
Computed from the prerequisite graph, not assigned. How this works