Raptor 0.6: A Small Agent Model on a New Base
Raptor 0.5 shipped and then it looped. What we found, why we moved to Spark-X2.5-4B, and how the 3.7 GB Raptor 0.6 was built for the Osaurus harness.
Today we're releasing Raptor 0.6, an open 4B agent model fine-tuned on the Osaurus tool surface and quantized to 3.7 GB. It reads your folders and spreadsheets, checks documents against data, drafts the follow-up, and runs the steps in between, all on your Mac. It decodes at about 105 tokens per second on an M5 Max and stays resident on an 8 GB machine. It needs Osaurus 0.25.0 or newer.
What happened to Raptor 0.5
Raptor 0.5 shipped on September 2. Within a week we had recordings of it looping: the same tool call re-issued until the harness gave up, paragraphs repeated word for word, and replies made of nothing but exclamation marks.
We spent the following days changing one thing at a time and found two causes. The exclamation marks were ours. A silent fallback in our serving stack routed the model through the wrong code path, and that is fixed in Osaurus 0.25.0. The repeated tool calls were the model. Once one identical turn sat in its history, it copied its own previous reasoning and call almost word for word. The untuned base does the same thing. We could not train it out, so we changed the base.
A new base
Raptor 0.6 is built on Spark-X2.5-4B, a dense model with ordinary key/value attention and no recurrent state to corrupt. We measured the raw base first. Then we trained twelve rounds of a light, attention-only fine-tune, each from the base, and only added training data when a measured failure called for it. Every round was checked against the untuned base on the same test suites.
The result reasons by default, calls tools in a format Osaurus parses natively, and ships with the base model's own sampling settings. No repetition penalty.
On the model page
- What went wrong with 0.5, and how we proved it
- Why the first 0.6 preview, on a looped transformer, was frozen
- The training recipe and the data it used
- Quantization measured against the bf16 reference
- The limits we found
- Harness changes that shipped in Osaurus 0.25.0
- Downloads and citation
Read the full story on the Raptor 0.6 model page.