Introducing Hunter 4B & 1.7B, our first agentic-coding models.
Today we’re releasing Hunter 4B and Hunter 1.7B, the first models from Lalakai Labs. We think both are among the best you can get at their size, and both are small enough to run on a laptop.
Hunter 4B
Hunter 4B is the stronger of the two. It’s tuned for reasoning, agentic coding, tool use, and security analysis, and it writes out its thinking inside <thought> tags before it answers, so you can see how it got there. That makes it easier to trust (and to debug) on math, code, and multi-step problems.
The 4-bit version is about 2.5 GB and needs roughly 3.5 GB of memory at a 4K context. If you have the room, the FP16 version (about 8 GB file, around 10 GB of RAM or VRAM) gives you the best output quality. It’s in GGUF format, so it works with llama.cpp, Ollama, and LM Studio.
Hunter 1.7B
Hunter 1.7B is the smallest model we could ship without giving up on real work. It’s built for defensive security: triaging alerts, answering incident-response questions, and working out how an attack likely unfolded, all without reaching for a bigger model.
It was trained on sequences up to 1,024 tokens, so keep prompts short for now. It
loads with Hugging Face Transformers using
trust_remote_code=True.
How we trained them
Hunter 4B is a fine-tune of Qwen 3.5 4B, trained with QLoRA on about 180,000 curated examples. We didn’t use raw web scrapes. Everything was deduplicated and written with explicit reasoning traces. The mix was 40% complex reasoning, 25% agentic coding, 20% multi-step tool use, and 15% world knowledge, with security-focused reasoning woven through all of it. We then quantized it with an importance matrix so the weights that matter most for reasoning survive the compression.
Hunter 1.7B is a fine-tune of XHToken’s Spark-X2.5-1.7B, trained with LoRA on 2,500 curated cybersecurity examples covering causal reasoning, offensive security, and incident-response Q&A.
Thanks to the Qwen team and the Spark-X authors for the base models.
Try them
The quickest way to run Hunter 4B:
ollama run hf.co/Lalakai/Hunter-1-4B:Q4_K_MBoth models are in the Hunter 1 collection on Hugging Face, with setup instructions and example code on each model card: