Kev 4B adapter for llama.cpp
Kev 1.0 by Jared Palmer, split into side files so one Qwen3.5-4B-Base can both run Kev's decisions (LoRA on) and write free text (LoRA off). Used by the Turtle Soup iOS app.
| File | What it is |
|---|---|
kev-lora-f16.gguf |
Kev's LoRA, converted to GGUF (f16) for llama.cpp. Apply it to Qwen3.5-4B-Base.Q4_K_M.gguf from mradermacher/Qwen3.5-4B-Base-GGUF. |
kev-head.bin |
Kev's pointer head as raw weights, applied to the base's hidden states. |
Conversion: llama.cpp's LoRA converter can't apply Qwen3.5's DeltaNet value-head reorder to adapter tensors, so the reorder was applied to the PEFT adapter first. Matches the merged ggml-org/Kev-4B-GGUF within noise on TurtleBench.
License: Apache-2.0, same as Kev and Qwen3.5.