espetro/kev-0.8b-gguf

Kev 0.8B v1.0 — packed GGUF

Kev-0.8B is a Jev-style "System One" decision model: a Qwen3.5-0.8B backbone plus a trained pointer head that returns calibrated probabilities for typed questions instead of generating text.

Built from Kev v1.0 — jaredpalmer/kev-0.8b @ bf75a6a (LoRA adapter + head.pt) merged onto Qwen/Qwen3.5-0.8B-Base @ dc7cdfe2 with espetro/llama.cpp's tools/kev pipeline (kev_v10_merge.py -> convert_hf_to_gguf --no-mtp -> kev_head.py -> kev_pack.py). The pointer head and calibration temperature (2.3511) are baked in as dec.head_* tensors + kev.* metadata.

Files

file	size	use
kev-0.8b-q8_0.gguf	814 MB	recommended for inference
kev-0.8b-f16.gguf	1.5 GB	re-quantization source / zero-drift reference

Use with the Kev-enabled llama.cpp fork

llama-server -hf espetro/kev-0.8b-gguf:Q8_0
# -> serves TypeSafe /v1/systemone + /studio automatically

curl localhost:8080/v1/systemone -H 'content-type: application/json' -d '{
  "state": "Shoes arrived two weeks late and in the wrong size.",
  "questions": {
    "department": {"type": "choice", "instructions": "Which team handles this?",
                   "criteria": {"returns": "Exchanges, refunds", "shipping": "Delays, lost packages"}}
  }
}'

llama-decide -hf espetro/kev-0.8b-gguf:Q8_0 --json request.json

The files still load in stock llama.cpp as an ordinary Qwen3.5 model — the kev.* metadata and head tensors are simply ignored there.

In the browser

espetro.github.io/llama.cpp runs a smaller demo quant of this model in llama.cpp WASM — see espetro/kev-0.8b-demo-gguf. q8_0 also works (?model= param) but is a ~2x larger download.

Source & license

• Weights + head: Apache-2.0 — jaredpalmer/kev
• Base model: Qwen/Qwen3.5-0.8B-Base (Apache-2.0)
• Pack/quantize + runtime: espetro/llama.cpp kev branch (MIT)