NeoHorse-1-4B tunes Qwen3.5 for agentic work

TokenRhythm has released NeoHorse-1-4B, a compact language model fine-tuned from Alibaba's Qwen3.5-4B base. At 4 billion parameters, it targets a familiar sweet spot: small enough to run on modest hardware, but tuned specifically for agentic behavior rather than general chat.

The model is aimed at three overlapping tasks — tool use, code generation, and step-by-step reasoning. That framing puts it squarely in the growing category of small models designed to act as the engine inside automated workflows, where a model calls external functions and chains multiple actions together instead of simply answering a single prompt.

Why it matters

Agentic capability has largely been the domain of larger frontier models, so the appeal of a 4B option is practical:

• Lower memory and compute requirements for local or edge deployment
• Faster inference for latency-sensitive tool-calling loops
• A permissive-enough footprint to experiment without frontier-scale infrastructure

A few caveats are worth noting. The repository lists the license simply as "other," so teams should check the terms before commercial use, and no context-length or benchmark figures accompany this initial release. As a first version from a smaller publisher building atop Qwen, NeoHorse-1-4B is best treated as a candidate to evaluate against established small agentic models rather than a proven drop-in. The full details are on its Hugging Face page.