Developer Reference
The models we customise into specialised language models, and the platforms they deploy to.
From Base to Specialised
A base model ships with broad general capability and no knowledge of your standards. We post-train it against verification signal built from your experts, until it performs your work to your bar. On open weights the result is a specialised language model you own and run; for Azure-hosted models the same customisation runs through Microsoft Foundry.
Supported Families
Most engagements start from an open-weight base we post-train with reinforcement learning. On Azure, we also customise hosted models through Microsoft Foundry. Smaller bases are cheaper to train and faster to serve, so we pick the smallest model that clears your quality bar.
Google's open-weight family, from a workstation-grade flagship down to phones.
Model card ↗Alibaba's open-weight family, from a 27B mid-size model down to 0.8B.
Model collection ↗OpenAI's open-weight family under Apache 2.0. Both sizes are sparse mixture-of-experts (MoE) models.
Model collection ↗Hosted models customised through Microsoft Foundry, plus Microsoft's own MAI line.
Fine-tuning docs ↗Deployment
Training runs and the finished model deploy inside your environment, on the platform your stack already trusts.
AWS
Modal
Azure
Databricks
SnowflakeWe support these, and more on request.
Tell us your domain, your constraints, and where your data lives. We will recommend the base model and deployment that fit.