TL;DR: "Open weights" means you can download the trained parameters and run them yourself. What you may do with them depends on the licence, and licences vary widely: OpenAI's gpt-oss, Qwen3.5 and DeepSeek-V4 ship under permissive licences (Apache 2.0 or MIT), while Meta's Llama 4 licence requires a separate licence above 700 million monthly active users, Kimi K2's modified MIT licence requires visible attribution above 100 million users or $20 million in monthly revenue, and some Mistral models need a commercial licence above $20 million in monthly revenue. Choose an open model on four checks: the licence fits your business, the model passes your evaluation, you can serve it at your volume, and you can trust where the weights came from.
Why teams choose open weights
Five reasons recur, and each one is a product requirement, not a preference.
- Data control. Prompts never leave your infrastructure, which settles residency and retention questions an API cannot.
- Pinning. The model does not change unless you change it, so behaviour stays stable between your releases.
- Customisation. Full fine-tuning, distillation and inspection of the weights.
- Cost at volume. At high, steady traffic, owned serving can beat API prices, as self-host vs API works out.
- Sovereignty. Governments and regulated sectors increasingly want models they can host locally; India's Sarvam AI, for example, released Sarvam-105B under Apache 2.0.
If none of these applies, an API is simpler to run, and many open models are also sold as hosted endpoints, which gives you their price without the operations.
The licence picture, November 2026
Each row is taken from the model card or licence file published by the developer. Read the full licence before relying on it; this is not legal advice.
| Model | Licence | Notable condition | Parameters (total / active) |
|---|---|---|---|
| gpt-oss-120b and 20b (OpenAI) | Apache 2.0 | None beyond Apache terms | 117B / 5.1B and 21B / 3.6B |
| Qwen3.5-397B-A17B (Alibaba) | Apache 2.0 | None beyond Apache terms | 397B / 17B |
| DeepSeek-V4-Pro and Flash | MIT | None beyond MIT terms | 1.6T / 49B and 284B / 13B |
| Llama 4 (Meta) | Llama 4 Community License | Above 700 million monthly active users, a licence Meta may grant "in its sole discretion" | Varies by model |
| Kimi K2 (Moonshot AI) | Modified MIT | Display "Kimi K2" above 100 million monthly users or $20 million monthly revenue | Varies by model |
| Mistral open models | Mostly Apache 2.0 | Some under a modified MIT requiring a commercial licence above $20 million monthly revenue | Varies by model |
Two observations. Permissive licences now cover some of the strongest open models, which lowers the legal barrier. And the strongest model from a family is not always open: Alibaba's Qwen4-Max (November 2026) is hosted only, with smaller open Qwen4 models promised later. Plan for the open tier to trail the hosted frontier.
The four checks
Licence. Read the use restrictions as well as the thresholds; some licences carry acceptable-use policies that bar categories of use. Note what changes between versions: a family can move from a permissive licence to a custom one in its next release, so pin the version you approved.
Evaluation. Run your golden set at the precision you will deploy, since quantization changes results, and compare against the API you would otherwise use.
Serving. Total parameters set the memory you must rent; for a mixture-of-experts model, that floor can be many accelerators even when compute per token is small. A smaller dense model that passes the same evaluation may be the cheaper choice.
Supply chain. Download weights only from the developer's official organisation, verify checksums, scan any custom loading code before running it, and keep a record of exactly which files you deployed. The OWASP 2026 list for LLM applications keeps supply chain among its top risks for good reason: a model file is executable risk if its loader runs code.
Policy beyond the licence
Some enterprises and governments restrict models by country of origin or require a review before adopting one. Check your organisation's procurement policy before choosing among technically equal candidates, and record the decision. Export rules and sanctions can also change what you may deploy where.
In the interview
Open-weight questions appear as "should we use an open model instead of the API?", "which open model would you build on?" or, at a lab, "should we release our weights?". A weak answer treats "open" as free and unrestricted. A Hire answer compares cost and quality. A Strong Hire answer reads the specific licence thresholds and use restrictions, tests on the team's own cases at deployment precision, works out the memory floor, verifies provenance, pins the version, and keeps an API as the fallback.
Practise with open-weight model or API and release our in-house model as open weights.
Next: self-host vs API and mixture of experts.