Qwen3.8-27B

EXL3 3 bpw · sglang · 200K context · RTX 3090 24 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.

139 tok/sdecode

583 tok/sprefill

200Kcontext window

Tested on this cardSep 25, 2026

Run it

With Omarchy Local AI it is one button. By hand, it is three steps: the weights, the config, the container. The server then answers on http://localhost:8000/v1.

1 Download the weights

hf download turboderp/Qwen3.8-27B-exl3 \
  --revision 6fe61ad620abfe97c5b49f9722c2bceeea4ccc28 \
  --local-dir ~/models/Qwen3.8-27B-exl3-6fe61ad6

2 Start the server

docker run --rm \
  --gpus all \
  -p 8000:30000 \
  --shm-size 16g \
  -e CUDA_DEVICE_ORDER=PCI_BUS_ID \
  -e HF_HUB_OFFLINE=1 \
  -e SGLANG_EXL3_EMBED_HOST=1 \
  -e SGLANG_EXL3_KERNEL=auto \
  -e SGLANG_EXL3_MODEL_PATH=/models/turboderp-Qwen3.8-27B-exl3-3.00bpw \
  -v ~/models/Qwen3.8-27B-exl3-6fe61ad6:/models/turboderp-Qwen3.8-27B-exl3-3.00bpw:ro \
  --entrypoint /opt/entrypoint.sh \
  ghcr.io/0xsero/sglang-exl3@sha256:84f75f3424c99a3d63392a4e0348292bdeba9cf7efba2ff1aa9b85c8fc0131e8 \
  python3 \
  -m sglang.launch_server \
  --model-path /models/turboderp-Qwen3.8-27B-exl3-3.00bpw \
  --quantization exl3 \
  --trust-remote-code \
  --host 0.0.0.0 \
  --port 30000 \
  --served-model-name turboderp-Qwen3.8-27B-exl3-3.00bpw \
  --context-length 204800 \
  --mem-fraction-static 0.80 \
  --kv-cache-dtype fp8_e4m3 \
  --mamba-ssm-dtype bfloat16 \
  --max-running-requests 1 \
  --max-mamba-cache-size 5 \
  --cuda-graph-max-bs-decode 1 \
  --chunked-prefill-size 1024 \
  --max-prefill-tokens 1024 \
  --reasoning-parser qwen3 \
  --tool-call-parser qwen3_coder \
  --speculative-algorithm NEXTN \
  --speculative-num-steps 3 \
  --speculative-eagle-topk 1 \
  --speculative-num-draft-tokens 4 \
  --speculative-token-map /opt/sglang-exl3/tokenmaps/qwen38_hot32k_v2.pt

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 RTX 3090 on Sep 25, 2026.

Details

Weights: turboderp/Qwen3.8-27B-exl3 @ 6fe61ad620 ›
Image: ghcr.io/0xsero/sglang-exl3@sha256:84f75f3424c99a3d63392a4e0348292bdeba9cf7efba2ff1aa9b85c8fc0131e8
Engine profile: sglang-qwen3.8-27b-exl3-3bpw-200k ›
Recipe file: registry/recipes/nvidia/rtx-3090-24gb/qwen3.8-27b.sglang.200k.json ›
Model released: Aug 5, 2026
Tested: vast, RTX 3090