AI
Trained in the Birdhouse.
Sized to the job.
Robin, Heron and Raven: three models built to execute, made to operate with Birdhouse OS.
The flock
Three models, one family
Our smallest model
Robinv0.7
Quick, light, everywhere.
Built for the high volume, low latency decisions that make up most of an agent's day. Small enough to run close to your data, cheap enough to call on every event.
Best for
- Classification and routing
- Extraction from documents
- On-device and edge work
Our general execution model
Heronv0.6
Steady and precise.
The everyday workhorse. It writes and reviews code, drafts and summarises, and follows a structured task through several steps, without the cost of reaching for the largest model.
Best for
- Writing and reviewing code
- Drafting and summarising
- Most agent workflow steps
Our smartest model
Ravenv0.6
The one that works it out.
For the hardest work: writing complex code, finding complex security vulnerabilities, and multi-step problems where getting it right matters more than getting it fast.
Best for
- Complex coding tasks
- Finding security vulnerabilities
- Reasoning and planning
Benchmarks
The flock, by the numbers
How each model scores across the work it does.
| Benchmark (SWE-bench) | Heron v0.6 |
|---|---|
| Coding | 87.0 |
| Reasoning | 68.2 |
| Tool use | 79.6 |
| Long context | 84.1 |
| Task Execution | 89.1 |
How they work together
The right model for every job
Birdhouse OS sends each task to the lightest model able to handle it, and moves up the family only when the work calls for it. Robin and Heron carry most of the load, Raven takes on the hard problems, and third-party models join if the task requires an extra hand.