---

license: apache-2.0

base_model: Qwen/Qwen3.6-27B

tags:

- gguf

- quantization

- shard-store

---

Qwen3.6-27B per-tensor quant shard store

A per-(tensor, type) quantization shard store for Qwen3.6-27B: every 2-D

weight tensor of the model, quantized independently into 40 GGUF types across

both the mainline llama.cpp and ik_llama.cpp families, plus a lossless BF16

reference shard per tensor. Assemble any mixed-precision GGUF from it without

re-encoding (see shards/index.json for the (tensor × type) grid).

Built directly from the official Unsloth BF16 conversion

(unsloth/Qwen3.6-27B-GGUF, included byte-identical under BF16/).

Layout

BF16/                          Unsloth BF16 source (2-part split, byte-identical)
imatrix_unsloth.gguf_file      Unsloth imatrix used for every imatrix-consuming encode
shards/
  main.shard                   header KV + non-quantizable tensors (norms, …)
  index.json                   { tensor -> {type -> path, offset, bytes} }
  MANIFEST.json                full provenance (sources, sha256s, encoder commits)
  token_embd/<TYPE>.shard
  output/<TYPE>.shard
  blk/<L>/<tensor>/<TYPE>.shard

Types (40)

• mainline (encoded by ggml-org/llama.cpp 571d0d5): Q4_0, Q5_0, Q8_0,

Q2_K–Q6_K, IQ1_S/M, IQ2_XXS/XS/S, IQ3_XXS/S, IQ4_XS/NL, MXFP4, NVFP4,

TQ1_0, TQ2_0, and the prism-ml Bonsai formats Q1_0 (binary, 1.125 bpw)

and Q2_0 (ternary, 2.25 bpw)

• ik_llama (encoded by ikawrakow/ik_llama.cpp 9d07d86): IQ2_K/KS/KL,

IQ3_K/KS, IQ4_K/KS/KSS, IQ5_K/KS, IQ6_K, IQ2_KT/IQ3_KT/IQ4_KT/IQ1_KT

(trellis), Q6_0

• BF16 lossless reference per tensor

Shared formats carry mainline-encoder bytes; ik encodes only its exclusives.

The Unsloth imatrix is applied to every type that accepts one. The

imatrix-required low-bit types (IQ1_S/M, IQ2_XXS/XS/S, IQ1_KT/IQ2_KT) are

omitted on the two uncalibrated tensors (token_embd, output).

Provenance

See shards/MANIFEST.json for exact source revisions, sha256s, encoder

commits, and toolkit state. Encoded with [qalloc]'s quantize producer

(dual-backend ggml dispatch, one shim per fork).