Open-weight models, on your own GPU.
isobata uses no external AI API. Every model runs on your server, next to your code. The weights are downloaded once, at installation, from their publishers, and they are all under the Apache 2.0 licence.
The models
| Model | What it does | Publisher | Licence |
|---|---|---|---|
| Qwen3 14B | Writes the answers, citing each source | Alibaba Qwen | Apache 2.0 |
| Qwen2.5-Coder 14B | Understands the question and routes it; writes descriptions of files and designs while indexing | Alibaba Qwen | Apache 2.0 |
| Qwen3-Embedding 0.6B | Turns code, docs, designs and questions into vectors for search | Alibaba Qwen | Apache 2.0 |
| BGE Reranker v2 M3 | Re-orders the search results by how well they answer the question | BAAI | Apache 2.0 |
All four run together on one NVIDIA GPU with 24 GB of memory. We choose and tune the models against a test set of real questions about real codebases; a change of model reaches you as a signed release, never silently.
How an answer is made
When indexing
- isobata reads your repositories, Markdown docs and Figma files and splits them into pieces: functions, sections, frames.
- It writes a short description of each file and design, so a question in plain words finds code that never uses those words.
- It turns every piece into a vector and stores it in the index on your server.
When your team asks
- The question is understood and, if it is broad, split into narrower questions.
- The index is searched for each of them, and the best candidates are re-ranked.
- The answer model writes the answer from those pieces only, and cites the file, page or frame behind each part, so anyone can check it.
Questions and documents don't need to share a language: a question in Italian finds documentation written in English, and the other way round.
What the models never do
- Learn from your code. The weights are fixed. Nothing is trained or fine-tuned on your code, your questions or the answers.
- Call out. The model servers sit on a private network with no route to the internet. See Security and architecture.
- Answer without sources. Every answer lists what it is based on.
What you need to run them: Installation.
Apply for the pilot