Palmyra x6

Palmyra x6 Technical Report: An Agentic, Tool-Use Model Post-Trained via Anchored Supervised Fine-Tuning

Peng Du, Kiran Kamble, Rakshith Vasudev, Zhizhuo Yang, Rohith Nadimpally, Arjun Krishna, Waseem AlShikh, Daniel M. Bikel

FIG. 01 Collect verified trajectories → Reject leaky rollouts → Fine-tune with a KL anchor → Train briefly and gently
Identifier
arXiv:2608.16620
Published
Aug 17, 2026
Authors
8
Type
Model
Area
Enterprise-optimized models

In one paragraph

Palmyra x6 is WRITER's model for enterprise agentic work. It was post-trained from a Mixture-of-Experts base on just 626 verified tool-use trajectories, with a KL anchor that keeps it close to the base model. It scored highest on BFCL Core in its comparison cohort.

BFCL Core, the highest in its comparison cohort
0.785
BFCL Core, the highest in its comparison cohort
verified synthetic tool-use trajectories used for post-training
626
verified synthetic tool-use trajectories used for post-training
epoch, at a low learning rate, with a KL anchor to the frozen base
1
epoch, at a low learning rate, with a KL anchor to the frozen base

FIG. 02: Mechanism

How it works

  1. Collect verified trajectories

    Synthetic tool-use rollouts span six task types, including financial research, coding, MCP tool suites and RAG. Each one is graded three times by an LLM judge, and a majority score of at least 0.5 is needed to pass.

  2. Reject leaky rollouts

    A rollout is dropped if it makes no tool calls, cites the demonstration, or shares more than 80% of its 4-grams with the teacher's reasoning. 626 trajectories make the final set.

  3. Fine-tune with a KL anchor

    The loss adds K times the KL divergence from the frozen base model, with K = 0.1. Any move away from the base costs extra, so the model learns tool use without drifting far.

  4. Train briefly and gently

    One epoch at a learning rate of 5×10⁻⁷ with global batch 16, using Muon on weight matrices and Adam on everything else.

FIG. 03: Interactive figure

Pull toward tools, anchor to base

Change the KL anchor strength and training length to see the trade between learning tool use and drifting from the base, then filter sample rollouts the way the training set was built.

Loading interactive demo…

FIG. 04: Notes

Summary

Palmyra x6 is built for the agentic, tool-heavy tasks enterprises run in Writer Agent. Instead of a large post-training run, it uses Anchored Supervised Fine-Tuning:

  • A Mixture-of-Experts base model (GLM-5.2, 744B total and about 40B active parameters) is fine-tuned on a small corpus of 626 verified synthetic tool-use trajectories.
  • Training runs for a single epoch at a low learning rate with a Muon and Adam hybrid optimizer.
  • A KL anchor to the frozen base model limits drift, so the model gains tool-use skill without losing general capability.

Results

  • 0.785 on BFCL Core, the highest score in its comparison cohort.
  • The highest six-benchmark mean in that cohort.
  • Substantial gains over the previous default model in Writer Agent, with competitive bias and safety results.

Palmyra X6 is one of the six models in our Harness Effect study.

Reference

Cite this work
@misc{du2026palmyrax6,
  title={Palmyra x6 Technical Report: An Agentic, Tool-Use Model Post-Trained via Anchored Supervised Fine-Tuning},
  author={Peng Du and Kiran Kamble and Rakshith Vasudev and Zhizhuo Yang and Rohith Nadimpally and Arjun Krishna and Waseem AlShikh and Daniel M. Bikel},
  year={2026},
  eprint={2608.16620},
  archivePrefix={arXiv},
  primaryClass={cs.CL}
}