Palmyra x6: an agentic, tool-use model
Palmyra x6 is WRITER’s shipped agentic model, built for multi-step planning, tool calling and long tasks in tool-rich environments such as MCP tool suites, web search and code sandboxes. Ordinary supervised fine-tuning on a small, specialized dataset tends to pull a model away from what it already does well, so the report uses anchored supervised fine-tuning: each token’s loss is weighted by the model’s confidence in that token, and a penalty keeps the model close to a frozen copy of the base.
I built the synthetic tool-use data pipeline whose data post-trained the model. A successful reference trajectory is reduced to a strategy-level demonstration with its answer removed, and the model then solves the task itself against live tools, using the demonstration only as guidance. A rollout is kept only if a judge model grades the whole trajectory correct, and a leakage filter rejects rollouts that copied the demonstration instead of calling the tools.
Main results
- Palmyra x6 has the top BFCL Core score among the compared models, at 0.785.
- It also has the top mean across the six benchmarks.
- It scores higher than the previous default model of Writer Agent on every benchmark in the evaluation suite.