Kev is a family of compact decision models built directly on Qwen3.5 and released on GitHub. The project appeared on Hacker News where it accumulated 172 points and 74 comments within days.
What Kev Actually Does
Kev models specialize in structured decision tasks rather than open-ended chat. They take input context and output discrete choices or rankings that follow explicit rules. The architecture stays close to the Qwen3.5 base while adding lightweight decision heads that reduce token waste on reasoning traces.
Size and Performance Numbers
No official parameter counts or benchmark tables were published in the initial release. Early users report the smallest variant fits in under 8 GB VRAM and runs at roughly 40 tokens per second on an RTX 4060. Larger variants in the family remain under 14 B parameters according to community measurements.
How to Try Kev Locally
Clone the repository and install with the provided requirements file. Load any checkpoint through the included Hugging Face-compatible script. A single inference call accepts a JSON context object and returns a structured decision plus optional score.
git clone https://github.com/jaredpalmer/kev
pip install -r requirements.txt
python infer.py --model kev-3b --input context.json
Pros and Cons
- Runs on consumer GPUs without quantization tricks
- Outputs are deterministic and easy to parse
- Fine-tuning data remains undisclosed
- No public leaderboard scores yet
- Limited to decision-style tasks only
Alternatives and Direct Comparisons
| Model | Base Size | Decision Focus | VRAM (approx) | License |
|---|---|---|---|---|
| Kev (small) | Qwen3.5 | Yes | <8 GB | Unknown |
| Phi-3 mini | 3.8 B | No | 6 GB | MIT |
| Gemma-2 9B | 9 B | No | 18 GB | Gemma |
Kev trades general capability for tighter decision formatting. Phi-3 mini offers broader language performance but requires extra prompting to match Kev's structured output.
Who Should Use Kev
Developers building rule-based agents or scoring systems gain the most. Teams that already run Qwen3.5 locally can swap in Kev without changing infrastructure. Skip Kev if your workload needs open-ended generation or multi-turn conversation.
Bottom Line
Kev shows that narrow decision models distilled from Qwen3.5 can deliver usable local performance with minimal hardware. The project remains early, yet the HN discussion indicates clear demand for this exact capability.
Further releases will likely add evaluation numbers and fine-tuning recipes.
Top comments (0)