---

license: apache-2.0

base_model: Qwen/Qwen3.8-27B

tags:

• gguf
• quantization

---

Qwen3.8-27B — per-type GGUF store (qalloc store-v2)

This repo is a byte-addressable quantization store, not a single quant.

It holds Qwen3.8-27B quantized at ~41 different types (mainline llama.cpp

and ik_llama.cpp families unified; per-type backend flag in the manifest):

• types/<TYPE>.gguf — the whole model at one type. Every file is a normal,

directly runnable uniform quant (a mainline-typed file runs on stock

llama.cpp; ik types need ik_llama.cpp).

• main.gguf — header KV metadata + all non-quantizable tensors (with the

upstream chat-template fix applied).

• manifest.json (qalloc-manifest/v2) — the byte index: for every tensor

× type, the exact [offset, offset+bytes) range inside its per-type file.

A client can assemble any mixed-precision allocation with plain HTTP

Range requests — no server-side work.

• calib/, eval/ — the calibration and evaluation corpus files used by

the damage-model program that drives allocation.

Assembly tooling and the allocation picker (exact MCKP solver over a

measured damage model) live in the qalloc project; the interactive picker

runs in the browser via WASM. The former v1 stores (per-tensor files, two

repos) are deprecated: this repo IS the store.