Custom Foundational Models. Model behavior.

We work with the latest open-weight releases from families like Qwen, Kimi, and GLM, and we choose the specific version per project based on your accuracy, latency, context-length, and hardware needs. Because the weights are open, you own the fine-tuned result and can run it on your own infrastructure instead of depending on a closed API. These families also ship vision-language variants, so we can use one toolchain whether your task is text-only or needs to read images and documents.

Supervised fine-tuning (SFT) updates the model's weights on your labeled examples and is the most direct way to lift accuracy on your domain. A LoRA adapter trains a small set of extra parameters that sit on top of the base model, which is faster and cheaper, lets you keep separate adapters for separate tasks, and can be merged into the base weights once it performs well. We often start with LoRA to iterate quickly, then commit to full SFT or merge the adapter for the production build.

Yes. We size the context window to your workflow so the model can read an entire contract, claim history, or knowledge base in a single pass instead of losing detail across chunks. For visual work we fine-tune vision-language models (VLMs) that take images, scans, screenshots, or video frames as input, so the same model can, for example, read a photo of a damaged part or a scanned invoice and respond in your terminology.

The data requirements depend on the approach. Fine-tuning a pre-trained foundation model often requires as few as 500 to 5,000 high-quality examples to achieve excellent results, since the base model already understands general patterns. Training a model from scratch typically requires tens of thousands of examples. During our discovery phase, we assess your available data and recommend the best approach. If your data is limited, we can supplement it with synthetic data generation or transfer learning techniques.

A typical custom model project takes 4-8 weeks from kickoff to production deployment. The first 1-2 weeks focus on data preparation and exploration. Training and validation usually take 1-3 weeks depending on model complexity. The final phase includes integration, testing, and deployment. Fine-tuning projects on existing foundation models are often faster, sometimes as quick as 2-3 weeks. We provide regular progress updates and intermediate results throughout the process.

Model development is inherently iterative, and we set clear performance benchmarks at the start of every project. If initial results fall short, we have multiple strategies to improve accuracy: collecting additional training data, engineering better features, trying alternative model architectures, or adjusting the problem framing. Our discovery phase includes a feasibility assessment so we can identify potential accuracy challenges before committing to full development. We do not consider a project complete until the model meets agreed-upon performance criteria.