Lalakai/Hunter-2-4B

Hunter 2 4B

Hunter 2 4B is a 4-billion-parameter conversational reasoning model from Lalakai Labs, built for local inference on consumer hardware. It works through multi-step findings inside <think>...</think> traces before giving its final answer, and is tuned for defensive security analysis, vulnerability triage, and agentic coding.

Hunter 2 4B is built on Spark-X2.5-4B from XHToken.

Weights are distributed in GGUF format for llama.cpp and Ollama. GGUF stores tensors only and does not use pickled code.

Highlights

• 4B parameters, optimized for defensive security, code review, and agentic coding
• Visible reasoning in <think> traces for multi-step findings
• Q5_K_M runs in about 3.4 GB of working memory at an 8K context, and the model was built for 8 GB laptops
• Apache-2.0 license

Model details

Developer	Lalakai Labs
Model type	Causal language model (conversational), GGUF
Parameters	4B
Context	4,096 tokens
Language	English
Version	v2.0
License	Apache-2.0

Training data

Hunter 2 4B was trained on approximately 100 million curated tokens, split between defensive-security scenarios and agentic-coding trajectories.

Category	Share	Focus
Defensive security	67%	Vulnerability remediation, cybersecurity evaluation, NIST guidance, cyber reasoning
Agentic coding	33%	Agentic coding trajectories and general coding work

Security data drew on vulnerability-remediation pairs, cybersecurity evaluation sets, NIST training material, and cyber-reasoning traces. Coding data came from agentic-coding trajectories. Roughly a quarter of the mix is clean-code grounding, so the model is trained to confirm benign code as safe and explain why rather than assume every audit finds a flaw.

Training method

Parameter-efficient QLoRA fine-tuning (rank 32, alpha 64) on layers 8–35, with the first eight layers, embeddings, and language-model head frozen. Training ran for 763 steps in BF16 at a batch of 2 with 16-step gradient accumulation, a cosine learning rate of 1e-4, and 5% warmup. It finished at a training loss of 0.729 (low 0.675) with no loss-spike breaker trips.

The weights were then quantized to Q5_K_M and Q8_0 with llama.cpp.

Files

File	Size	Description
hunter2-Q5_K_M.gguf	~2.98 GB	5-bit quantized weights. Recommended for most users.
hunter2-Q8_0.gguf	~4.38 GB	8-bit quantized weights. The high-fidelity reference build, with minimal quantization loss.

Quick start

llama.cpp

Run inference from the CLI:

llama-cli -m hunter2-Q5_K_M.gguf --temp 0.2 -p "Audit this function for command injection flaws."

To serve over HTTP:

llama-server -m hunter2-Q5_K_M.gguf --port 8080 -c 4096

Python (llama-cpp-python)

from llama_cpp import Llama

llm = Llama(
    model_path="hunter2-Q5_K_M.gguf",
    n_ctx=4096,
    verbose=False,
)
print(llm("Audit this function for command injection flaws.", max_tokens=256)["choices"][0]["text"])

Ollama

Create a Modelfile. The GGUF carries its own chat template, so only the runtime parameters and system prompt need to be set:

FROM ./hunter2-Q5_K_M.gguf

PARAMETER temperature 0.2
PARAMETER top_p 0.90
PARAMETER top_k 40
PARAMETER repeat_penalty 1.1

SYSTEM """You are Hunter 2, an autonomous security analysis agent developed by Lalakai Labs. Confirm safe code as safe and state why; deliberate only on genuinely multi-step findings."""

ollama create hunter2 -f Modelfile
ollama run hunter2

Hugging Face CLI

hf download Lalakai/Hunter-2-4B hunter2-Q5_K_M.gguf

Choosing a file

• Local security review on a typical laptop: Q5_K_M
• Highest output quality, if you have the memory: Q8_0

Intended use

• Defensive security analysis, vulnerability triage, and remediation guidance
• Secure-code review and multi-step code reasoning
• Agentic coding on local hardware
• Security-oriented reasoning, with human oversight

Evaluation

No formal benchmark results have been published for this model yet. They will be added to this card when available.

Limitations

• A 4B model will not match frontier-scale models on the hardest reasoning tasks or on niche knowledge. Use the Q8_0 weights when quality matters most.
• The model was trained on 4,096-token sequences. Chunk long files and long alert histories. The Spark-X2.5 base supports a much longer native window, but quality beyond the trained length is not guaranteed.
• Like all language models, Hunter 2 4B can produce hallucinations, reflect biases in its training data, and generate unsafe content when prompted adversarially. Review outputs for your use case and apply your own safety filtering in production.
• It is trained to confirm clean code as safe, so a "no findings" answer is not proof that code is secure. Treat its output as one input to a review, not the review itself.

Responsible use

Use this model under human supervision and within applicable law and your organization's policies. Do not use it to target systems you are not authorized to test, and do not use it to generate weaponized exploits or turnkey attack infrastructure.

License

Released under the Apache License, Version 2.0.

Contact

Built by Lalakai Labs.

Acknowledgements

Hunter 2 4B is a fine-tune of Spark-X2.5-4B (Apache-2.0). We thank the XHToken team for the base model.