---
license: apache-2.0
base_model: ggml-org/Kev-4B
tags: [gguf, decision-model, chat, llama.cpp]
---

# bend-4b — one model, two modes

GGUF quants of **Kev-4B** (Apache-2.0, upstream: [ggml-org/Kev-4B-GGUF](https://huggingface.co/ggml-org/Kev-4B-GGUF)) — the **4B model** of the Bend shelf.

**One model, two modes:** a single loaded copy serves **chat** and **structured decisions**; switching modes **swaps heads on the same model** — never a second model, never a reload. See the head shelf: **[bend-heads](https://huggingface.co/OmenWorks/bend-heads)**.

## Files & provenance (stated changes)
- `Kev-4B-Q8_0.gguf` — Q8_0, 4.48 GB — **byte-identical mirror** of the upstream file (sha256 `7c2ebed90560522c…`, matches ggml-org/Kev-4B-GGUF). No changes.
- `Kev-4B-Q4_K_M.gguf` — Q4_K_M, 3.03 GB — **byte-identical mirror** of upstream's own Q4_K_M (sha256 `33ae6b18926502b2…`, matches ggml-org/Kev-4B-GGUF). No changes. Its decision head `cls.output` is Q8_0 and its embedding Q4_K: the shape the in-browser (WebGPU) decider reads.
  - *Replaced 2026-10-17:* this path previously held our own 2.71 GB requant (sha256 `93c66545eff17854…`), whose `cls.output` had been requantized to Q4_K; the in-browser decider refuses that head, so every decision failed. That file remains in this repo's git history.

## License
Apache-2.0 — full text in `LICENSE` (this repo). Upstream copyright/license notice retained; upstream name credited above and in the filenames. No endorsement by the upstream authors is implied.