Qwen3.8-27B

· llama.cpp · 128K context · Radeon RX 7900 XT · 20 GB

Alibaba's 27B dense model from August 2026, and the strongest model that fits one consumer card. It thinks before it answers, calls tools reliably and reads images, which makes it the default for coding agents and long projects.

43 tok/sdecode

472 tok/sprefill

128Kcontext window

Tested on this cardOct 5, 2026

Run it

Follow the steps below to download and run this recipe. Omarchy Local AI lists the recipes it can launch on your hardware. The server then answers on http://localhost:8080/v1.

1 Set up first

# Not Docker; Linux on an RX 7900 XT (gfx1100, 20 GB). The archive is llama.cpp b11160 (70c4e1582, unmodified source) built with the ROCm 10.0.0 toolchain, GPU_TARGETS=gfx1100 and GGML_CUDA_FA_QUANTS including q8_0-q5_1 (stock builds lack the K q8_0 / V q5_1 FlashAttention kernel; BUILD.txt inside). It does not bundle the ROCm runtime: install ROCm 7.x system libraries (tested with 7.2.4 in /opt/rocm/lib: libamdhip64, hipblas, rocblas, hsa-runtime). Keep other GPU applications light: the 128K window is sized for the 20 GB card, with roughly 1 GiB left for the desktop.
# Full instructions: https://github.com/victoralcazardev/rx7900xtx-local-llm/tree/511bf1b

2 Install the engine

curl -L -o llama-b11160-rocm10-gfx1100-kvmix-linux-x64.tar.gz https://github.com/victoralcazardev/rx7900xtx-local-llm/releases/download/engine-b11160-rocm10-gfx1100-kvmix/llama-b11160-rocm10-gfx1100-kvmix-linux-x64.tar.gz
echo 'f1cb8e2683c94cc5891a11af7876058d50e2c0556f0ffeb92d2451bf659fc3ec  llama-b11160-rocm10-gfx1100-kvmix-linux-x64.tar.gz' | shasum -a 256 -c
tar xzf llama-b11160-rocm10-gfx1100-kvmix-linux-x64.tar.gz

3 Download the weights

hf download ISTA-DASLab/Qwen3.8-27B-GSQ-RCO-GGUF Qwen3.8-27B-GSQ-RCO-IQ3_S-mtp.gguf \
  --revision d562806dbafae37109975e970aae91b43e73b440 \
  --local-dir models/Qwen3.8-27B-GSQ-RCO-GGUF-d562806d

4 Start the server

llama-b11160-rocm10-gfx1100-kvmix/llama-server \
  -m models/Qwen3.8-27B-GSQ-RCO-GGUF-d562806d/Qwen3.8-27B-GSQ-RCO-IQ3_S-mtp.gguf \
  -c 131072 \
  -ctk q8_0 \
  -ctv q5_1 \
  -fa on \
  -np 1 \
  --ctx-checkpoints 4 \
  -ngl all \
  -ub 256 \
  --spec-type draft-mtp \
  --spec-draft-n-max 3 \
  --temp 1.0 \
  --top-p 0.95 \
  --top-k 20 \
  --min-p 0.0 \
  --reasoning-effort medium \
  --alias qwen3.8-27b \
  --host 127.0.0.1 \
  --port 8080

What it passed

✓ loadLoads and serves within an hour

✓ chatAnswers a plain question and stops on its own

✓ reasoningThinks separately and gets 17 × 23 right

✓ toolsCalls a tool with the right arguments and uses the result

✓ contextRecalls a code buried in a prompt that fills 85% of the window

✓ speedDecodes at 15 tok/s or more

Passed all six checks on a real AMD Radeon RX 7900 XT on Oct 5, 2026.

Details

Weights: ISTA-DASLab/Qwen3.8-27B-GSQ-RCO-GGUF @ d562806dba ›
Image: none: a program on the host
Source: victoralcazardev/rx7900xtx-local-llm @ 511bf1b ›
Launch: llama.cpp-rocm-qwen3.8-27b-gsq-rco-iq3-s-mtp-128k ›
Recipe file: registry/recipes/amd/rx-7900-xt-20gb/qwen3.8-27b.llama.cpp.128k.json ›
Model released: Aug 5, 2026
Tested: owner, AMD Radeon RX 7900 XT