What is open-weight AI, and what can you do with the weights?
An open-weight model is one whose weights — the billions of numbers it learned in training — are published for anyone to download and run, usually on Hugging Face and under a licence. You get the model itself, not the data or the code that produced it, which is why open-weight is narrower than open source. The weights are what make a choice of operator possible: the same DeepSeek or Qwen model can answer from Paris or from China, and the company that runs it — not the lab that trained it — decides which law can reach your prompts.
Published 6 Oct 2026 · checked on 20 Oct 2026
What the weights are
Training adjusts a neural network's parameters until its outputs match its training data well. Those final values are the weights: a set of files, published with a model card that gives the size, the licence and the intended uses. Load the files into an inference engine and you have the model — the same one the maker benchmarked. Quantisation stores the same weights at lower precision, 8 or 4 bits instead of 16, to save memory, at a cost in quality that you have to measure on your own tasks.
Size is counted in parameters. Many recent models are mixtures of experts, with two numbers: the total, which must all sit in memory, and the active share used for each token, which sets most of the compute per token. Three examples from Evanesca's catalogue: Qwen3.5 397B, 397B (17B active); DeepSeek V4 Flash, 284B (13B active); GPT-OSS 120B, 117B (5.1B active).
What you get, and what you don't
| Open-weight model | Closed model | |
|---|---|---|
| The weights | Published, downloadable | Never published |
| Where it can run | On any operator's servers, or yours | Only through the maker and the partners it picks |
| Training data and code | Rarely published | Not published |
| Terms of use | A licence, from permissive to restrictive | The maker's terms of service |
| Who decides where your prompts go | The operator you choose | The maker |
What you do not get is the recipe. Open source in the strict sense — the Open Source Initiative's definition — also requires the training code and enough information about the data to rebuild an equivalent model; very few models meet it. What open-source AI means sets out the difference.
Read the licence before the scores
Published weights are not free of conditions. Apache-2.0 and MIT, the permissive licences of many open models, allow commercial use. Others add terms: the Llama 3.3 Community License requires a licence from Meta above 700 million monthly active users, and Qwen's own licence sets its threshold at 100 million. A few releases forbid commercial use altogether.
Each model page on Evanesca states its licence and links the maker's card; the licence table in our open-source guide lists them side by side.
How much memory the weights need
A rough rule gives the order of magnitude: parameters × bytes per parameter. At 16-bit precision that is two bytes, at 8-bit one, at 4-bit half a byte. The memory for the context — the KV cache — comes on top, and grows with the length of conversations and the number of users.
| Model | Parameters | 16-bit | 8-bit | 4-bit |
|---|---|---|---|---|
| Mistral Small 3.2 | 24B | 48 GB | 24 GB | 12 GB |
| Llama 3.3 70B | 70B | 140 GB | 70 GB | 35 GB |
| Qwen3 235B | 235B (22B active) | 470 GB | 235 GB | 118 GB |
| Qwen3.5 397B | 397B (17B active) | 794 GB | 397 GB | 199 GB |
| GLM-5.2 | 753B | 1,506 GB | 753 GB | 377 GB |
| Kimi K3 | 2.8T (104B active) | 5,600 GB | 2,800 GB | 1,400 GB |
An 80 GB data-centre GPU such as the NVIDIA H100 holds Llama 3.3 70B only at 8-bit or below, with little room left for context; models of several hundred billion parameters need a multi-GPU server. For scale: on 6 October 2026, OVHcloud listed an instance with one H100 at €2.80 an hour before VAT, about €2,044 for a 730-hour month, busy or idle (prices).
Who runs the weights decides the jurisdiction
Open weights can run in four kinds of places, and each puts your prompts under different law:
- Your own machines: nobody else holds the data, at the cost of buying and running the hardware.
- An operator under EU control, such as OVHcloud AI Endpoints or Scaleway Generative APIs, in France.
- The EU region of a US cloud: the servers are in Europe, but a US company must disclose data in its “possession, custody, or control” wherever it is stored (18 U.S.C. § 2713).
- The maker's own service: DeepSeek's app, for one, stores personal data in China according to its privacy policy.
On Evanesca, every model runs in Europe — most of them at Scaleway in Paris, with OVHcloud in Gravelines as a fallback — and Evanesca keeps no copy of your prompts; Scaleway may keep requests that cause errors or look abusive for up to 14 days. The exception to open weights there is Mistral Large 4, in preview: its maker, Mistral AI, runs it in its own EU region and keeps inputs and outputs 30 days to monitor abuse. Evanesca serves no model outside Europe. See sovereignty for the full route.
Weights are frozen; models move on
A published set of weights never changes. Improvements arrive as new releases under new names — DeepSeek V4 Flash and DeepSeek V4.1 Flash are two generations, not one model updated — so name the exact version you rely on, and re-test when you switch. DeepSeek V4.1 Flash shows how far two releases of one family can differ. Operators retire versions too, so a service built on one set of weights has to plan its move to the next.
The same holds for answers: on Evanesca, the passport under each answer names the model that produced it, so a result can be traced to a specific set of weights and a specific operator.
Trying open-weight models without a GPU
You do not need hardware to compare open-weight models: an operator's API does it per token. Evanesca's chat and OpenAI-compatible API run them in Europe, most of them in France, from €0.168 per million input tokens with Gemma 4 26B, VAT included, with credit from €2, and the exact cost of each answer shows under it. Evanesca is not open yet; this site shows how it will work.
Start with the models, put two side by side with compare, and check each price on pricing; open-weight models in 2026 suggests which to use for what.
Frequently asked questions
What is an open-weight model?
A model whose trained parameters — its weights — are published for anyone to download and run, usually under a licence. Mistral Small 3.2, DeepSeek V4 Flash and Llama 3.3 70B are examples. The training data and code usually stay private.
Is open-weight the same as open source?
No. Open source, as the Open Source Initiative defines it, also requires the code used to train and run the model and enough information about its training data to rebuild an equivalent system. Open-weight models publish the weights only.
Can I run an open-weight model on my own computer?
Small ones, yes, especially quantised: Mistral Small 3.2 takes about 12 GB at 4-bit. Large ones need data-centre GPUs: Llama 3.3 70B needs about 140 GB at 16-bit, and GLM-5.2 about 753 GB at 8-bit, before any memory for context.
Who sees my prompts when I use an open-weight model?
The company that runs it. If a French operator runs it in France, your prompts stay with that operator; if the maker's own service runs it, they go to the maker. On Evanesca, the open-weight models run at OVHcloud, Scaleway, AKI.IO or Infomaniak, and Evanesca keeps no copy of your prompts; Scaleway may keep requests that cause errors or look abusive for up to 14 days.
Read next
- What is open-source AI?Open-source AI lets anyone use, study, modify and share a model, with its weights, code and training-data information. What qualifies, and what it changes.
- French open-source AI: the companies, the models, and where they runMistral AI, Hugging Face, OVHcloud, Scaleway: who does what in French open-source AI, and why a model's origin does not decide where your data goes.
- DeepSeek V4.1 Flash: what it is, and where it runsDeepSeek V4.1 Flash: a 552B multimodal MoE with a 1M context, under MIT, released on 10 September 2026. On Evanesca, AKI.IO runs it in Germany.