Qwen3.8-27B-pi

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

bytkim's Qwen3.8-27B fine-tune for the Pi coding-agent harness, trained with supervised fine-tuning and reinforcement learning. Supports low, medium and xhigh reasoning; the EXL3 build retains a 4-bit MTP head.

132 tok/sdecode

816 tok/sprefill

200Kcontext 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:8000/v1.

1 Download the weights

hf download mrq/Qwen3.8-27B-pi-exl3-3.00bpw \
  --revision 3412e8c5a6cafdc75924fd52e339dee72de88285 \
  --local-dir ~/models/Qwen3.8-27B-pi-exl3-3.00bpw-3412e8c5

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 PYTORCH_CUDA_ALLOC_CONF=expandable_segments:True \
  -e SGLANG_EXL3_EMBED_HOST=1 \
  -e SGLANG_EXL3_KERNEL=auto \
  -e SGLANG_EXL3_MODEL_PATH=/models/bytkim-Qwen3.8-27B-pi-exl3-3.00bpw \
  -e SGLANG_EXL3_PREFILL_FP16_ACC=1 \
  -e SGLANG_EXL3_VIT_MLP_CHUNK=8192 \
  -e SGLANG_EXL3_VIT_SDPA=1 \
  -e SGLANG_OPT_MAMBA_SKIP_DECODE_LOCK=1 \
  -e SGLANG_VLM_CACHE_SIZE_MB=256 \
  -v ~/models/Qwen3.8-27B-pi-exl3-3.00bpw-3412e8c5:/models/bytkim-Qwen3.8-27B-pi-exl3-3.00bpw:ro \
  --entrypoint /opt/entrypoint.sh \
  ghcr.io/0xsero/sglang-exl3@sha256:c8922bd7256caf1b273af7ac49221e652be9e6072c1cdec4b959f506fd415d79 \
  python3 \
  -m sglang.launch_server \
  --model-path /models/bytkim-Qwen3.8-27B-pi-exl3-3.00bpw \
  --quantization exl3 \
  --trust-remote-code \
  --host 0.0.0.0 \
  --port 30000 \
  --served-model-name bytkim-Qwen3.8-27B-pi-exl3-3.00bpw \
  --context-length 204800 \
  --mem-fraction-static 0.82 \
  --kv-cache-dtype fp8_e4m3 \
  --mamba-ssm-dtype bfloat16 \
  --max-running-requests 1 \
  --max-mamba-cache-size 3 \
  --cuda-graph-max-bs-decode 1 \
  --cuda-graph-max-bs-prefill 240 \
  --chunked-prefill-size 4096 \
  --max-prefill-tokens 4096 \
  --prefill-decode-interval 16 \
  --reasoning-parser qwen3 \
  --tool-call-parser qwen3_coder \
  --mamba-radix-cache-strategy no_buffer \
  --disable-overlap-schedule \
  --image-processor-backend pil \
  --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 Oct 5, 2026.

Details

Weights: mrq/Qwen3.8-27B-pi-exl3-3.00bpw @ 3412e8c5a6 ›
Image: ghcr.io/0xsero/sglang-exl3@sha256:c8922bd7256caf1b273af7ac49221e652be9e6072c1cdec4b959f506fd415d79
Launch: sglang-qwen3.8-27b-pi-exl3-3bpw-mtp-vision-200k ›
Recipe file: registry/recipes/nvidia/rtx-3090-24gb/qwen3.8-27b-pi.sglang.200k.json ›
Model released: Sep 29, 2026
Tested: owner, RTX 3090