🤖 AI

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

Palmyra x6 is an enterprise-oriented agentic language model that achieves state-of-the-art performance on tool-use benchmarks through a conservative, controlled post-training approach using Anchored Supervised Fine-Tuning on a compact, verified dataset.

Original authors: Peng Du, Kiran Kamble, Rakshith Vasudev, Zhizhuo Yang, Rohith Nadimpally, Arjun Krishna, Waseem Alshikh, Daniel M. Bikel

Published 2026-08-18

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Peng Du, Kiran Kamble, Rakshith Vasudev, Zhizhuo Yang, Rohith Nadimpally, Arjun Krishna, Waseem Alshikh, Daniel M. Bikel

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

In the rapidly evolving landscape of artificial intelligence, a persistent challenge has been how to teach large language models to act as reliable assistants without breaking the very skills they already possess. Traditionally, creating a high-performing model for a specific job required rebuilding the entire system from scratch, a process akin to constructing a new engine for every different vehicle. More recently, researchers have discovered that it is often sufficient to take an existing, powerful model and carefully refine its behavior for a specific purpose, provided the training data is of exceptional quality and the learning method is precise. This approach focuses on "agentic" tasks, where the model must not just answer questions but actively plan steps, use digital tools like web search or code execution, and navigate complex, multi-step workflows to solve problems. The goal is to create an AI that can function effectively in a real-world digital environment, handling tasks that require long-term planning and interaction with various software systems, rather than simply generating text.

A team of researchers at Writer, Inc. has addressed this challenge with the development of Palmyra x6, a model designed specifically to excel at these agentic tasks. Rather than training a massive new model from the ground up, they started with a sophisticated existing foundation known as GLM-5.2, a model with hundreds of billions of parameters. To this foundation, they applied a highly specialized training method called Anchored Supervised Fine-Tuning. This technique is designed to teach the model new skills—specifically, how to use tools and plan complex tasks—while strictly preventing it from forgetting or degrading the general knowledge and reasoning abilities it already possessed. The result is a system that can navigate intricate digital environments, such as those involving web searches, code execution, and document retrieval, with a level of competence that significantly outperforms previous versions of their own agent and competes with the most advanced models available.

The core of this achievement lies in the quality and nature of the data used for training. Instead of feeding the model vast amounts of internet text, the researchers generated a compact, highly curated dataset consisting of just 626 examples. Each example is a complete, verified trajectory of a task being solved by an AI agent. These are synthetic stories where the AI plans a series of steps, calls various tools to gather information or perform actions, and arrives at a correct solution. To ensure these examples were perfect, the researchers used a rigorous process where a "teacher" model first demonstrated a successful path, and a "student" model was then asked to solve the same problem independently, using the teacher's plan only as a high-level guide. If the student model attempted to bypass the process by copying the answer or failed to use the tools correctly, the attempt was discarded. Only the most successful, honest attempts were kept, creating a small but incredibly high-quality library of examples that taught the model how to think and act like a competent agent.

To learn from this small dataset without losing its original capabilities, the researchers employed a method that acts as a safety anchor. During the training process, the model is constantly reminded of its original, untrained self. This "anchor" ensures that while the model learns to use tools and plan tasks, it does not drift away from the general intelligence and safety behaviors it started with. This is crucial because training on a tiny dataset could otherwise cause the model to become overly specialized or lose its ability to handle general conversation. By keeping the model tethered to its original state, the researchers were able to inject new agentic skills without eroding the foundation. They also utilized a specialized mathematical optimizer, a tool that helps the model learn more efficiently by treating its internal connections as geometric shapes rather than just numbers, allowing it to make better use of the limited data available.

The results of this approach were immediate and substantial. When tested against a wide range of benchmarks designed to measure an AI's ability to use tools, plan long-term tasks, and follow complex instructions, Palmyra x6 demonstrated a massive improvement over the previous default model used by Writer's agent. It showed particularly strong gains in financial research, where it had to search the web and analyze data, and in general tool-use scenarios. The model also performed competitively against other leading frontier models, often ranking at the top of the group in average scores across multiple different tests. Perhaps most importantly, the model maintained a high degree of safety and neutrality. In tests designed to measure political bias and refusal behavior, Palmyra x6 showed a balanced approach, presenting multiple sides of controversial issues and refusing harmful requests at rates consistent with its base model, proving that the new skills did not come at the cost of safety or reliability.

The researchers were careful to note that this model is a specialized tool, not a universal replacement for all AI tasks. Its strengths are specifically in agentic workflows where it can call tools and execute plans, while its performance on other types of tasks remains governed by the capabilities of the base model it was built upon. The training data, while small in number, was sufficient to teach the specific behavior of tool use because the learning method was designed to be conservative and precise. The model is now available for commercial use, offering a version that can be run efficiently on standard hardware. This work demonstrates that with the right combination of high-quality data, a careful training objective, and a stable foundation, it is possible to create highly capable AI agents without the need for massive, resource-intensive retraining from scratch. The success of Palmyra x6 suggests a future where AI assistants can be tailored to complex, real-world jobs with precision and safety, relying on a few hundred perfect examples rather than millions of imperfect ones.

Technical Summary: Palmyra x6

Problem Statement

The primary challenge addressed by this work is the efficient development of large language models (LLMs) capable of agentic task completion—specifically multi-step planning, tool/function calling, and long-horizon execution within complex environments (e.g., MCP suites, web search, code execution, and retrieval-augmented pipelines).

Historically, achieving high performance in such specialized domains required retraining the entire pipeline from pretraining through post-training. The authors posit that for high-performing base models, it often suffices to apply targeted post-training, provided the data is high-quality and the objective is carefully chosen to avoid eroding the base model's existing capabilities. A specific technical hurdle identified is the risk of capability degradation during standard Supervised Fine-Tuning (SFT), where training on external expert demonstrations can push the model's distribution far from its own, causing it to lose general abilities.

Methodology

1. Architecture and Base Model

Palmyra x6 is built upon the GLM-5.2 base model, inheriting the GlmMoeDsa architecture unchanged. Key architectural features include:

Scale: 744B total parameters with ~40B active parameters per token (Mixture-of-Experts).
Attention: Multi-head Latent Attention (MLA) and DeepSeek Sparse Attention (DSA) with IndexShare (cross-layer reuse of sparse attention indices), requiring specific inference stack support.
Context: 65,536 tokens.

2. Data Synthesis and Curation

The training corpus consists entirely of synthetic agentic trajectories (626 curated examples).

Source: Task prompts were sourced from internal subject matter experts and external benchmarks.
Generation Pipeline: The authors employ a Self-Distillation Fine-Tuning (SDFT) adapted for tool use. A "teacher" model generates a reference trajectory, which is sanitized (removing specific values and answers but retaining strategy and tool order) and injected into the "student" model's system prompt. The student then executes the task using live tools.
Filtering: A rigorous multi-stage gate ensures quality:
- Structural Filters: Enforce tool usage, cap conversation length (≤50 messages), and limit tool call repetition/errors.
- Verification: A judge model scores the student's trajectory against the teacher's reference answer.
- Leakage Control: Trajectories are rejected if the student solves the task by reading the demonstration rather than using tools (e.g., high n-gram overlap with the teacher's reasoning).
Scale: The dataset is intentionally small (~626 trajectories) to adhere to the "less is more" principle, relying on high-quality, verified long-horizon examples.

3. Training Objective: Anchored Supervised Fine-Tuning (ASFT)

To prevent the model from drifting away from its base capabilities while learning new agentic skills, the authors utilize Anchored Supervised Fine-Tuning (ASFT). The loss function combines two components:

DFT Token-Probability Weighting: Weights the negative log-likelihood of tokens by the model's own detached probability, focusing learning on tokens the model is already confident about.
KL Anchor: A penalty term ( $K \cdot KL(\pi_{ref} \parallel \pi_{\theta})$ ) against a frozen copy of the base model ( $\pi_{ref}$ ). This constrains the distributional drift. The authors selected a KL weight of K=0.1.

4. Optimizer: Muon + Adam Hybrid

The training run replaces the standard Adam optimizer with a hybrid approach:

Muon: Applied to 2-D weight matrices (attention projections, FFN weights). Muon treats weight matrices as geometric objects, orthogonalizing momentum buffers via Newton–Schulz iterations to equalize update singular values. This improves scaling efficiency and stability.
Adam: Retained for 1-D parameters (embeddings, biases, router weights) where orthogonalization is undefined.
Configuration: Global batch size of 16, single epoch, learning rate of $5 \times 10^{-7}$ with cosine decay.

Key Contributions

Recipe for Efficient Agentic Specialization: Demonstrates that a 744B parameter model can be effectively specialized for complex tool-use tasks using a compact corpus of ~626 synthetic trajectories, avoiding the need for massive datasets or full pretraining.
Anchored Fine-Tuning Strategy: Validates the use of ASFT with a KL anchor ( $K=0.1$ ) to add agentic skills without eroding the base model's general capabilities or alignment.
Muon Optimization at Scale: Successfully applies the Muon optimizer to a 744B MoE model, leveraging its geometric properties for more efficient updates on 2-D matrices while maintaining stability.
Rigorous Synthetic Data Pipeline: Establishes a robust pipeline for generating, sanitizing, and filtering synthetic tool-use trajectories that prevent reliance on demonstrations and ensure on-policy learning.

Results

Internal Benchmarks: Palmyra x6 shows substantial improvements over the previous default Writer Agent model across six evaluation benchmarks. Notable gains include +0.320 on MCP-Atlas, +0.305 on FinanceBench, and +0.304 on IFBench.
External Comparisons: Against five frontier models, Palmyra x6 achieved the highest mean score (0.765) across the six benchmarks and topped the BFCL Core score at 0.785.
Safety and Bias:
- On the Washington Post ModelSlant evaluation, Palmyra x6 presented both sides of political arguments 80% of the time, the highest rate among tested models.
- It demonstrated strong safety performance on the FORTRESS adversarial benchmark, with the system message configuration achieving an adversarial safety score of 67.0 (vs. 58.4 for the base model) with negligible cost to benign helpfulness.
- The model showed no statistically robust quality asymmetry across opposing political positions in English.

Significance and Claims

The paper positions Palmyra x6 as a case study in targeted behavioral specialization. The authors claim that by combining a high-quality, small-scale synthetic dataset with a conservative training recipe (ASFT with a strong KL anchor and the Muon optimizer), it is possible to create a highly capable agentic model that retains the robustness and alignment of its base model.

The significance lies in the efficiency and control of the approach:

It challenges the notion that large-scale data is strictly necessary for complex agentic behaviors.
It demonstrates that "less is more" when the data is verified and the training objective is anchored to the base distribution.
It emphasizes that safety and alignment are preserved not just by the data, but by the training dynamics (KL anchor) and the system-level deployment (guardrails, sandboxing), rather than by re-aligning the model from scratch.

The authors explicitly state that public benchmarks serve only as a sanity check; the primary validation comes from internal datasets mimicking enterprise use cases, where the model shows superior performance. The model is released under a commercial license with BF16 weights and an FP4 variant for memory-efficient serving.