Qwythos-9B v3 Released: Optimizing Chat Templates for Agentic Reasoning

EmperoAI has released version 3 of Qwythos-9B, specifically addressing critical failures in agentic harnesses by refining preserved and adaptive thinking mechanisms within the model's chat template.

Addressing Agentic Harness Performance

The release of Qwythos-9B v3 comes as a direct response to observed instabilities in agentic workflows. According to the developers, previous iterations exhibited significant issues when deployed in agentic harnesses, where the model's ability to maintain consistent reasoning chains was compromised.

Improvements in Adaptive Thinking

The core of the v3 update focuses on the chat template, specifically targeting "preserved and adaptive thinking." In the context of Large Language Models (LLMs), these mechanisms are crucial for the model's ability to maintain a coherent internal monologue or "Chain of Thought" (CoT) while adapting its strategy based on environmental feedback—a requirement for successful autonomous agent behavior.

The developers describe the improvement as a "night and day difference," suggesting that the updated template significantly enhances the model's reliability and reasoning capabilities during complex task execution.

Deployment and Availability

To implement these fixes, users are advised to redownload the model weights. The update is available in both GGUF (for quantized deployment via llama.cpp) and Safetensors (for full-precision or HF-based loading) formats.

Note: Due to the brevity of the source announcement, specific architectural changes beyond the chat template adjustments were not disclosed.