GitHub radar
NeoHorse-1-4B: Agentic Fine-tune of Qwen3.5-4B
TokenRhythm has done a routing-guided agentic post-training on Qwen3.5-4B from Alibaba, that’s +10 points on agent benchmarks and can be run on your laptop!
We’ve applied routing-guided agentic post-training (a special kind of training technique to improve performance on agentic use cases such as tool use, task planning, and coding) to Alibaba’s Qwen3.5-4B, which achieved a +5.93 macro-average improvement on 10 benchmarks compared to Qwen3.5-4B baseline and showed remarkable improvements on agentic tasks (+9.79 on WorkBuddy Bench, +6.21 on QwenClawBench, +9.75 on HumanEval). The original context window size is 262144 tokens. The 4-bit version of the model (4.2B params) takes about 2.5 GB disk space. You can try out this model via GGUF repack by mradermacher here: https://huggingface.co/mradermacher/NeoHorse-1-4B-GGUF on LM Studio or Ollama.
Why a vibe-coder should care
You can run the model locally for simple agent tasks such as repeating scripts, simple code generation and automations without internet. It’s a better starting point than the untrained Qwen3.5-4B if you care about simple agent tasks. Note that it’s still 4B params so there are obvious differences to state-of-the-art models like Claude, but it’s great for some simple repetitive tasks or when you don’t have the budget to pay for a subscription.
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