Open-weight AI models in 2026: which to use for what, and who runs them
Open-weight models now handle most everyday AI work, from long documents to code and images, but the label alone tells you little: not the licence, not the hardware, not who runs the model when you call it. This guide sorts the models Evanesca serves by what they are good for, with size, licence, price and operator taken from its catalogue. Every model on it runs in Europe, mostly in France: Evanesca serves no model outside Europe. Mistral Large 4, which Evanesca also runs in Europe, is left out: no weights published yet.
Published 6 Oct 2026 · checked on 20 Oct 2026
What “open-weight” tells you, and what it does not
An open-weight model is one whose trained weights are published, so anyone with the hardware — you, or a cloud provider — can run it. It is not quite open source: the training data and code usually stay private. Three things still have to be checked one by one.
- The licence: what you may do with the weights and the outputs. Some are permissive (Apache-2.0, MIT); others are the maker's own terms, with conditions.
- The size: total parameters, and for a mixture-of-experts model the smaller number active for each token, which drives speed and price.
- The operator: when you call a model through an API, the company that runs it decides where your prompt is processed and which law can reach it.
Evanesca's catalogue prints all three for every model, along with the result of its six checks. The tables below are built from it, so a price or a context window that changes there changes here. Quotes from model cards were read on 6 October 2026.
Large generalists for long, demanding work
For long documents, analysis, planning and hard code, the largest open-weight models in Evanesca's catalogue are DeepSeek V4.1 Flash, GLM-5.3, DeepSeek V4 Flash, GLM-5.2, Qwen3.5 397B and Qwen3 235B. DeepSeek V4 Flash, Qwen3.5 397B and Qwen3 235B are mixture-of-experts designs: Qwen3.5 397B has 397B parameters but uses only 17B for each token, so a token costs far less compute than the total size suggests.
| Model | Maker | Parameters | Context | Reasoning | Price in · out | Run by |
|---|---|---|---|---|---|---|
| DeepSeek V4.1 Flash | DeepSeek | 763B | 1M | Yes | €1.008 · €3.024 | AKI.IO, Germany (fallback: Scaleway) |
| GLM-5.3 | Zhipu / Z.ai | 753B | 512k | Yes | €3.024 · €9.24 | AKI.IO, Germany (fallback: Scaleway) |
| DeepSeek V4 Flash | DeepSeek | 284B (13B active) | 256k | Yes | €0.672 · €1.344 | Scaleway, Paris |
| GLM-5.2 | Zhipu / Z.ai | 753B | 256k | Yes | €3.024 · €9.24 | Scaleway, Paris |
| Qwen3.5 397B | Qwen | 397B (17B active) | 250k | Yes | €1.008 · €6.048 | Scaleway, Paris |
| Qwen3 235B | Qwen | 235B (22B active) | 250k | — | €1.26 · €3.78 | Scaleway, Paris |
Their makers describe them in their own terms. Z.ai calls GLM-5.2 “our latest flagship model for long-horizon tasks” (model card). DeepSeek presents the 0731 release of DeepSeek V4 Flash as the official version, with “substantially enhanced agentic capabilities” (model card). Qwen says Qwen3.5 397B covers 201 languages and dialects and reads images as well as text (model card). Treat these as claims to test on your own prompts, not as a ranking.
Smaller models for everyday volume
Classification, extraction, short answers and high volume rarely need a giant. The rest of Evanesca's catalogue, sorted by output price:
| Model | Maker | Parameters | Context | Reasoning | Price in · out | Checks |
|---|---|---|---|---|---|---|
| Pixtral 12B | Mistral AI | 12B | 128k | — | €0.336 · €0.336 | 5/6 · Security ✕ |
| Mistral Small 3.2 | Mistral AI | 24B | — | — | €0.252 · €0.588 | 5/6 · Security ✕ |
| Gemma 4 26B | 26B (4B active) | 256k | — | €0.168 · €0.84 | 6/6 | |
| GPT-OSS 120B | OpenAI | 117B (5.1B active) | — | Yes | €0.252 · €1.008 | 5/6 · Security ✕ |
| Mistral Small 4 | Mistral AI | 119B (6.5B active) | 256k | — | €0.2772 · €1.1088 | 5/6 · Security ✕ |
| Qwen3 Coder 30B | Qwen | 30.5B (3.3B active) | 256k | — | €0.336 · €1.344 | 5/6 · Security ✕ |
| Llama 3.3 70B | Meta | 70B | — | — | €1.512 · €1.512 | 5/6 · Security ✕ |
| Qwen3.6 35B | Qwen | 36B (3B active) | 256k | Yes | €0.42 · €2.52 | 6/6 |
| Apertus v1.5 70B | Swiss AI Initiative | 72B | 100k | — | €1.2936 · €4.62 | 4/6 · Security ✕ · Benchmarks ✕ |
| Qwen3.8 27B | Qwen | 27.8B | 256k | Yes | €1.008 · €5.544 | 6/6 |
OpenAI released GPT-OSS 120B and its smaller sibling under Apache-2.0; its card says the larger one fits “into a single 80GB GPU” thanks to quantised weights (model card) — useful to know if you weigh running it yourself (the arithmetic is here). The cheapest input price in Evanesca's catalogue is €0.168 per million tokens, for Gemma 4 26B. A dash in the context column means the window has not been verified at the operator, never that there is none.
Models that read images
On Evanesca, DeepSeek V4.1 Flash, Qwen3.8 27B, Apertus v1.5 70B, Qwen3.6 35B, Gemma 4 26B, Mistral Small 4, Qwen3.5 397B, Mistral Small 3.2 and Pixtral 12B accept images as well as text. Use them for photographed documents, screenshots, charts and whiteboards: the image is read by the model in Europe, like the text.
Pixtral 12B, from Mistral AI, is the one built for images: its card says it was trained on images and text together and takes several images, of any size, in one message (model card). On Evanesca it reads screenshots and documents well but misreads an image made of a single flat colour, and it did not pass the security check — keep it for reading, not for content you do not trust. The other models above read images too, and all of them can call tools — why tool calling depends on the operator.
The European-made option
Most open-weight models are made outside Europe; what is European in Evanesca's catalogue is who runs them, and where. Some are made here — Apertus v1.5 70B, Mistral Small 4, Mistral Small 3.2 and Pixtral 12B — and two of them show the range. Mistral Small 3.2, from Mistral AI in Paris: 24B parameters, Apache-2.0, run by Scaleway, Paris (fallback: OVHcloud); its card tags two dozen languages, among them French, German, Spanish, Italian, Portuguese, Polish, Romanian and Swedish (model card). And Apertus v1.5 70B, from the Swiss AI Initiative: 72B parameters, Apache-2.0, run by Infomaniak, Switzerland (model card).
Mistral Small 3.2 has been run by Scaleway in Paris since 20 October 2026, with OVHcloud in Gravelines taking over if Scaleway fails. The wider French scene — makers, hosts, models — is covered in French open-source AI.
Licences at a glance
| Licence | Open-weight models under it | What to check |
|---|---|---|
| Apache-2.0 | Qwen3.8 27B, Apertus v1.5 70B, Qwen3.6 35B, Gemma 4 26B, Mistral Small 4, Qwen3.5 397B, GPT-OSS 120B, Qwen3 Coder 30B, Qwen3 235B, Mistral Small 3.2 and Pixtral 12B | Permissive, with an explicit patent grant |
| MIT | DeepSeek V4.1 Flash, DeepSeek V4 Flash and GLM-5.2 | Permissive, few conditions |
| GLM-5.3 License | GLM-5.3 | The maker's own terms: read them before commercial use |
| Llama 3.3 Community License | Llama 3.3 70B | Meta's terms: “Built with Llama” shown on products built on it |
Apache-2.0 and MIT both allow commercial use, modification and redistribution with light obligations; Apache-2.0 adds an explicit patent grant. Every other licence in the table is the maker's own text: read it before building a product on the model.
The six checks, model by model
Every model in Evanesca's catalogue shows six checks — licence, provenance, security, GDPR, benchmarks and a live test (how each check is run) — with the ones it passed and the ones it misses or has not had yet. Among Evanesca's open-weight models, DeepSeek V4.1 Flash, GLM-5.3, Qwen3.8 27B, DeepSeek V4 Flash, GLM-5.2, Qwen3.6 35B, Gemma 4 26B and Qwen3.5 397B pass all six.
Apertus v1.5 70B, Mistral Small 4, GPT-OSS 120B, Qwen3 Coder 30B, Qwen3 235B, Mistral Small 3.2, Llama 3.3 70B and Pixtral 12B miss the security stamp: they did not pass all of the injection, jailbreak, leak, exfiltration and harmful-request tests. They remain available, and the passport under each of their answers names the missing stamp.
Frequently asked questions
What is the difference between open-weight and open-source?
An open-weight model publishes its trained weights, so it can be run by anyone with the hardware. Open source in the strict sense would also publish the training code and data, which most makers keep private. What you may do with the weights depends on the licence, which ranges from permissive (Apache-2.0, MIT) to the maker's own terms.
Which open-weight model should I start with?
It depends on the job. For long, hard work, start with GLM-5.2, Qwen3.5 397B or DeepSeek V4 Flash; for cheap everyday volume, a small model such as Gemma 4 26B; for images, one of the multimodal models. Then test two or three on your own prompts and compare the exact cost shown under each answer.
Do my prompts go to China when I use a model made there?
Not on Evanesca: the models made in China are run in Europe — in France by OVHcloud or Scaleway, in Germany by AKI.IO — and their makers do not receive your prompts. How a model behaves is a separate question, which the security check answers.
Why do prices differ so much between models?
Size and the number of active parameters drive the operator's cost, and reasoning models also bill their thinking as output tokens. Evanesca's prices are per million tokens, VAT included, with no subscription; every model's price is on the pricing page.
Read next
- What is open-weight AI, and what can you do with the weights?An open-weight model publishes its weights, the numbers it learned in training. What you get, what you don't, how much memory they need, and who can run them.
- 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.