How we choose your base model
Every model we build starts from an open-weight base model. We choose that base so you can own, run and change the result without asking anyone’s permission.
Our rule
MIT or Apache 2.0 by default.
Both licences let you use, modify, fine-tune and redistribute the model commercially, with no user caps or revenue limits. What we build on them is yours to keep.
Some strong models carry extra conditions, such as revenue thresholds, naming rules or usage policies. We only use them when you ask, and the conditions are written into your assessment before any work starts. Nothing is used silently.
Current shortlist
Current shortlist.
Reviewed November 2026. The open-weight field moves fast, so we recheck this list every quarter and confirm the licence file of the exact checkpoint before every project.
| Model family | Developer | Origin | Licence | Typical fit |
|---|---|---|---|---|
| Gemma 4 | United States | Apache 2.0 | Small to mid-size models; multimodal; on-device and single-GPU | |
| gpt-oss | OpenAI | United States | Apache 2.0 | Reasoning; single-GPU to single-node |
| Mistral (Small 4, Large 3) | Mistral AI | France (EU) | Apache 2.0 | EU-origin requirement; mid to large |
| Qwen3.6 | Alibaba | China | Apache 2.0 | Strong all-rounder; single-GPU fine-tuning |
| DeepSeek V4 | DeepSeek | China | MIT | Frontier-class quality; multi-GPU to cluster |
| GLM-5 | Zhipu AI | China | MIT | Coding and agent workloads; multi-GPU to cluster |
Licence terms can differ between checkpoints in the same family. We check the licence file of the exact version we use.
Model origin
Some boards and risk committees restrict where a base model comes from.
Tell us at the start.
If origin is restricted, we choose from US and EU models (Gemma, gpt-oss, Mistral). The decision and the reason for it are recorded in your model card, so your second line and your regulator can see them.
If origin is not restricted, we compare candidates from every family on your own data and pick on measured results.
What you receive
For whichever base model we use:
- The upstream licence file and notices, as shipped with the weights
- A provenance record: model name, exact version, source repository and file checksums
- A model card stating the base model, its licence, its origin and why it was chosen
- Your fine-tuned weights and adapters, which are yours under the base licence
What we don’t promise
Some claims we won’t make.
- Legal advice. We tell you what the licence says and how we’ve applied it. Your counsel makes the final call.
- That today’s best model stays best. We can re-base your system on a newer model later; your test suite tells you whether it’s worth it.
Questions
What legal teams ask.
Can you use a model that isn’t on this list?
Yes, if your use case needs it. We’ll show you its licence conditions in the assessment, and we’ll only use it if you accept them in writing.