NeoHorse-1-4B: A 4B Model Built for AI Agents and Tool Use
Overview
NeoHorse-1-4B is a 4B-parameter causal language model from TokenRhythm, derived from Qwen3.5-4B and adapted through routing-guided agentic post-training for text-based agent harnesses, tool use, coding, and instruction following. Its key distinction is not a new model scale or modality, but the training process: a routing harness assigns tasks to a heterogeneous model pool, records tool interactions and outcomes, estimates capability demand, and uses capability-level feedback to shape later training mixtures. The model accepts text and produces text, supports a native 262,144-token context window with an advertised extension up to 1,010,000 tokens, and uses BF16 Safetensors weights. The release contains language-model weights only; it excludes vision weights and is repackaged for text-only inference. You run it with the Transformers ecosystem and can serve it through SGLang or vLLM using OpenAI-compatible APIs. The most important practical point is that the reported gains target agentic execution, coding, and instruction following under a...
Copyright of this story solely belongs to hackernoon.com. To see the full text click HERE
