PAW-27B-GGUF

PAW-27B

If this model is useful to you, support the work by following @nawariokra on X, or buy me a coffee on Ko-fi.

<a href="https://ko-fi.com/F4P3296XP7"><img src="https://ko-fi.com/img/githubbutton_sm.svg" alt="Support me on Ko-fi" height="32"></a>

A ~2.17 bpw trellis-coded quantization of Qwen/Qwen3.8-27B that runs on one 24 GB GPU with a 256k context at 38–78 tok/s.

Requirements — read this first

This is not a standard GGUF. It uses the PAW trellis format (paw-dense arch) and requires the llama-paw fork. Stock llama.cpp cannot read this model.

How it compares

Against two same-family IQ2XXS GGUFs, measured by us on the same machine with byte-identical harnesses:

PAW-27B leads MMLU-Pro, HumanEval+, MBPP+ and strict IFBench. Unsloth's file is 0.548 GB smaller and gives up 12.4 points of MMLU-Pro. AtomicChat is larger than PAW-27B and does not recover the coding gap.

Honest summary

These are peer comparisons, not retention numbers. No HumanEval, MBPP or GSM8K figure exists for Qwen/Qwen3.8-27B at any precision, so there is no published denominator to claim retention against on those axes, and we are not going to invent one. What the table shows is that at roughly the same file size this quantization is ahead of the two obvious alternatives — not how much of the unquantized parent survives.

The references that do exist for this parent are IFBench (79.5 unquantized, official Qwen) and LiveCodeBench v6 (90.3). The IFBench figure here is a 64-item defect sentinel, not the 300-item benchmark, so it is directional only and not comparable to the official 79.5.

Serving

See SERVING.md for the full configuration. Short version:

• requires llama-paw, not stock llama.cpp
• speculative decoding with the parent's own MTP head (--spec-type draft-mtp) — 22.7 → 78.4 tok/s, lossless
• -ub 2048 is a hard ceiling at -c 262144, not a preference
• KV at q8_0 is free; -fa on is required for it
• for agent/tool use, do not send reasoning back across turns — see the loop limitation below

Limitations

• Agent clients loop unless reasoning is dropped from history. The model copies its own prior <think> block verbatim and then repeats the action it went with. Serve with --no-reasoning-preserve and configure the client not to return reasoning_content. Details and measurements in SERVING.md.
• Unbounded thinking returns empty answers. Use --reasoning-budget 2048, not -1.
• Rare-glyph copy defect. The model cannot emit ➞ (U+279E) and aborts generation on prompts containing it; in this fork the server returns HTTP 500. Eight HumanEval prompts contain it and it reproduces deterministically under greedy decoding. This is a model defect, not a harness artifact, and the affected tasks are scored as failures in the numbers above.
• No parent denominator on HumanEval/MBPP/GSM8K (see above).
• IFBench here is a 64-item sentinel, not the full benchmark.
• Long-context quality is unmeasured; only speed and footprint were measured at 191k–256k.

Integrity

sha256 f5ec4b409de07f43bbe7bb5d9b3215181c53f5a26876821f732839a2ea3ab8cf  PAW-27B.gguf

License

Apache 2.0, inherited from the base model Qwen/Qwen3.8-27B.