Qwen3.8-27B-GGUF

Qwen3.8-27B · GGUF (imatrix)

Qwen3.8-27B for llama.cpp — importance-matrix quantized from the original BF16. Runs from 16 GB VRAM up.

The main weight files carry the language model. Vision ships as a separate mmproj file and the model's multi-token-prediction head as a separate speculative draft — llama.cpp keeps both outside the main file by design, not because anything was cut down. Take either, both, or neither.

Renamed 2026-08-21 from Qwen3.8-27B-Text-Only-GGUF. The old name described the layout of one file but read as "this model cannot see", which was never true — the vision projector has been in this repo since it was first published. Old links redirect automatically.

The file names still say Text-Only. Hugging Face redirects a renamed repo but not individual file URLs, so renaming the files would 404 every existing direct link and download script. They stay as they are.

Community quantization. Not an official Qwen release; not endorsed by or affiliated with the Qwen team or Alibaba Cloud. "Qwen3.8" identifies the upstream model this artifact derives from (Apache-2.0 §6).

---

Which file do I download?

All three columns measured on one RTX 3090 (24 GB), Vulkan, full offload. Context figures come from llama-fit-params, i.e. what actually fits — not a calculation.

The size/context trade is steeper than the size/quality trade. Dropping from Q6_K to Q4_K_M costs a little quality and buys 3.4× the context — on the same card.

The counter-intuitive part: IQ4_XS is 9 % smaller than Q4_K_M but ~23 % slower. IQ-family quants cost more compute to dequantize. Take IQ4_XS because you need the size, not because you want speed. If Q4_K_M fits your card, it is both faster and higher quality.

16 GB card: IQ4_XS. Q4_K_M technically loads but leaves almost nothing for KV. 24 GB card: Q4_K_M for speed, Q6_K for quality. Apple Silicon: unified memory is the budget — 32 GB → Q5_K_M/Q6_K, 64 GB → Q8_0.

Vulkan numbers. CUDA builds are typically faster.

---

Why the KV cache is unusually small

Qwen3.8 is a hybrid-attention model. Of its 64 layers only 16 are full attention — the other 48 are Gated DeltaNet linear attention and hold no KV cache. With num_key_value_heads = 4, head_dim = 256:

A comparable dense-attention 27B needs roughly four times this — which is why the context figures in the table above are as large as they are.

The table is measured on a 24 GB card. For a 16 GB card with IQ4_XS (14.0 GiB) the same arithmetic gives roughly 32K at f16 KV or 64K at q8 — estimated, not measured, as the author has no 16 GB card to test on.

---

Vision

llama-mtmd-cli -m Qwen3.8-27B-Text-Only-Q4_K_M.gguf \
               --mmproj mmproj-Qwen3.8-27B-f16.gguf \
               --image photo.jpg -p "Describe this image."

The projector is the general-purpose tower from Qwen3.8-27B, unchanged. Dedicated Qwen3-VL-* models will still do better on dense OCR and small-object counting.

---

Speculative decoding (MTP)

Qwen3.8 was trained with a multi-token-prediction head. llama.cpp's converter publishes the target model and the MTP draft as two files by design, so the head is here as its own optional download rather than inside the main weights.

llama-cli -m Qwen3.8-27B-Text-Only-Q4_K_M.gguf \
          -md mtp-Qwen3.8-27B-Q4_K_M.gguf \
          --spec-type draft-mtp --spec-draft-n-max 1 \
          -p "..."

Works with llama-cli and llama-server. llama-completion does not accept -md. Requires build b10502 or newer — draft-mtp does not exist before it.

The numbers

One RTX 3090, Q4_K_M target, 512 tokens, greedy, single stream. no draft and n-max 1 are the mean of three runs; the rest are single runs.

CUDA

+36 % at n-max 1. Every setting helps on CUDA, but the default of 3 leaves a third of the gain on the table.

The backend matters more than anything else here

The same card, same model, same prompt, on the Vulkan build:

+3 % at best, and negative at the default. Acceptance is the same 77 % — the draft is doing its job either way. What differs is the cost of batched verification, which Vulkan does not make cheap enough to pay for the draft's own forward pass.

Metal (M1 Max, 64 GB unified, same files, same method):

Negative. −11 % at n-max 1, −21 % at 2.

So: +36 % on CUDA, +3 % on Vulkan, −11 % on Metal — from identical files, at the same 75–77 % acceptance on all three. The draft does its job everywhere; what differs is what batched verification costs on each backend. On Apple Silicon, do not attach the draft. If you are on ROCm, SYCL, or anything else, measure before assuming any of these numbers applies to you.

Apple Silicon users: llama.cpp is not the fastest path for this model anyway. An MLX build of the same weights ran 15.9 tok/s on the same machine — 42 % faster than llama.cpp Metal, and no draft involved.

Only one draft file, and why

Q8_0 (2.9 GiB) was built and tested too: 55.0 tok/s at 73.8 % acceptance, versus 56.1 at 75.0 % for this Q4_K_M (1.9 GiB) — and the two produced byte-identical output. The bigger draft was slower and no more accurate, so it is not published. A draft cannot change what the target accepts, so there is no quality argument for it.

It is not bit-identical to running without a draft

Speculative decoding preserves the output distribution — every drafted token is verified against the target, so nothing gets through that the target would not have produced. It does not reproduce the same string. Batched verification changes floating-point reduction order, so a near-tie between two candidate tokens can land on the other one.

Measured: each configuration is perfectly reproducible with itself (two greedy runs byte-identical), but greedy output with the draft diverged from greedy output without it after ~338 characters on CUDA and ~121 on Vulkan, continuing as an equally coherent answer. If you need byte-reproducible output, do not attach a draft.

---

How these were made

Quantized from the original BF16 weights, not re-quantized from an existing INT4 release — so no compounding loss.

1. Vision tower and MTP block split off at the safetensors level; model.language_model.* promoted to model.*
2. convert_hf_to_gguf.py --outtype bf16 --no-mtp → BF16 GGUF (lossless from source). The mmproj and mtp-* files come from second passes over the unmodified source with --mmproj and --mtp — the pairing llama.cpp's converter documents
3. Importance matrix over 300 chunks of the same calibration corpus used for this author's W4A16 releases (512 passages, pile-val news text), on 2×RTX 3090
4. Every level quantized with that imatrix, K-quants included

Every file in this repo was loaded and generated with before publishing — including both mmproj files, which were checked against a synthetic image with known content. Not just checksum-verified.

---

Requirements

A llama.cpp build that knows the qwen35 architecture. Build b10502 or newer works.

Older builds fail on the main files with check_tensor_dims: tensor 'blk.64...' not found. That is a converter that counted an MTP block the main files do not carry — not a corrupt download. The mtp-* files are the opposite case: they are block 64, and --spec-type draft-mtp only exists in b10502 and newer.