GitHub radar
NeoHorse-1-4B: agentic fine-tune of Qwen3.5-4B
TokenRhythm fine-tuned Alibaba's Qwen3.5-4B with routing-guided agentic post-training — +10 points on agent benchmarks, runs on any laptop.
TokenRhythm fine-tuned Alibaba's Qwen3.5-4B using routing-guided agentic post-training — a training approach focused specifically on agentic scenarios: tool use, task planning, and code. The model gained +5.93 points on a ten-benchmark macro average over the base Qwen3.5-4B, with the sharpest gains on agentic work: +9.79 on WorkBuddy Bench, +6.21 on QwenClawBench, +9.75 on HumanEval. Native context window is 262,144 tokens. At 4.2B parameters, the 4-bit version occupies roughly 2.5 GB. A ready-to-use GGUF repack by mradermacher is available at https://huggingface.co/mradermacher/NeoHorse-1-4B-GGUF for LM Studio or Ollama.
Why a vibe-coder should care
If you want a local model for simple agentic tasks — repeating scripts, lightweight code generation, offline automation — this is a better-tuned starting point than the raw Qwen3.5-4B base. Honest caveat: 4B is 4B, and the quality gap compared to frontier models like Claude is real and significant. This fits the use case of simple repetitive tasks or situations where a paid subscription is not an option.
How to install
Copy this and send it to your agent — Claude Code, Codex, any of them:
Install NeoHorse-1-4B in GGUF format from https://huggingface.co/mradermacher/NeoHorse-1-4B-GGUF — pick a quantization level for my machine, set it up in LM Studio or Ollama, and show me how to send it a request.
A regular laptop with 8 GB of RAM — the 4-bit version takes about 2.5 GB.
Open on Hugging Face▌ More finds