Open-weight models are usually framed as a quality story, and that framing is out of date. Several are close enough in daily work that the deciding factors are elsewhere: whether you can run them yourself, what a token costs at your volume, and where the data goes.
This page is about choosing one for a reason rather than for a benchmark, using DeepSeek, Qwen, and Le Chat as the three common entry points.
The short answer
Pick DeepSeek when cost per token dominates and a hosted API is acceptable. Pick Qwen when you need an open-weight model you can self-host or fine-tune. Pick Le Chat when you want a European vendor with a generous free tier and a strong interface. None of them is the default daily driver for most teams, and that is fine.
What open weights change
| Property | Closed API | Open weights |
|---|---|---|
| Model version | Changes under you | Pinned until you upgrade |
| Data location | Vendor's servers | Your hardware, if you run it |
| Cost shape | Per token | GPU hours + ops |
| Fine-tuning | Usually limited | Yours to do |
| Support | Vendor | Community |
The first and second rows are the ones teams underestimate. A pinned model version means evaluations stay valid; data location is often a procurement requirement rather than a preference.
DeepSeek: the cost lever
DeepSeek's practical appeal is price per token, which for high-volume work is a different budget line than the frontier models. The hosted API makes it a drop-in choice for pipelines where the task is well specified and the model is not the interesting part.
The caveats are ordinary: hosted means vendor-hosted, with the retention and policy questions that implies. If the reason you wanted open weights was control, the hosted product does not give it to you.
Qwen: the deployment choice
Qwen offers a model family with open weights across sizes, including code and vision variants, plus a hosted chat. That breadth makes it the practical option for a team that wants the same model in a prototype and in a self-hosted deployment.
Running it yourself is an engineering project: GPU capacity, a serving stack, quantisation choices, and the evaluation that proves the smaller variant is good enough. The reward is a fixed model and a data path you control.
Le Chat: the product alternative
Le Chat is a consumer product first: fast interface, generous free tier, European hosting. Its model is competitive without being the benchmark leader, and the reason to use it is often the vendor relationship rather than the model.
For a European team with data-residency requirements and no appetite for self-hosting, that combination is the argument. It is a product decision, not a technical one, and should be evaluated as such.
The hidden cost of self-hosting
A self-hosted assistant is not one binary. It is a serving stack, a model, an evaluation harness, and a person responsible for upgrades. At low volume the arithmetic does not work; at high volume it often does, and the break-even is the thing to model.
The mistake is choosing self-hosting for ideological reasons and discovering the operational load afterwards. Decide who maintains it before deciding to run it.
Where each fits
High-volume, well-specified tasks: DeepSeek's hosted API is the cheapest path. Data that cannot leave your infrastructure: Qwen or another open-weight family, self-hosted. A European vendor with a strong consumer product: Le Chat.
For the general daily-driver question, the big closed assistants still win on integration and polish — see choosing a daily-driver assistant.
Evaluating one properly
Use your own twenty cases, not a public benchmark. Run them against the candidate and against your current assistant, score pass or fail, and decide from the difference. A model that is two points worse on a leaderboard but ten percent better on your tasks is the right choice.
Record the model version with the results. The version is the variable that invalidates the comparison, and it changes without notice on hosted APIs.
Frequently asked questions
Are open-weight models safe to use commercially? Read the licence attached to the weights; they differ. Some are permissive, some have usage restrictions, and the licence is not the same as a hosted product's terms.
Can I fine-tune one on company data? Yes, and that is a real advantage. Budget for the evaluation that proves the fine-tune helped, because it often does not.
Which is the best free option? Le Chat's free tier is generous for daily use; DeepSeek's hosted API is cheap but metered. Both are free to try, and that is the point. See general assistants.
Last checked: 10 November 2026.
Related: choosing a daily-driver assistant, context and cost, and general assistants.
