Qwen3.8-Flash-Next

EXL3 3.05 bpw · sglang · 256K context · RTX 5070 Ti · 16 GB

Qwen3.8's large mixture-of-experts model: far more knowledge than the 27B, with few active parameters, so it stays fast.

44 tok/sdecode

3,338 tok/sprefill

256Kcontext window

Tested on this cardOct 16, 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 turboderp/Qwen3.8-Flash-Next-exl3 \
  --revision 69e33439ae950f17bcbe95c98f117d80f759ab6d \
  --local-dir ~/models/Qwen3.8-Flash-Next-exl3-69e33439

2 Start the server

docker run --rm \
  --gpus all \
  -p 8000:30000 \
  --shm-size 16g \
  -e CUDA_DEVICE_ORDER=PCI_BUS_ID \
  -e EXL3_MOE_CPU_THREADS=8 \
  -e HF_HUB_OFFLINE=1 \
  -e PYTORCH_CUDA_ALLOC_CONF=expandable_segments:True \
  -e SGLANG_EXL3_CPU_TIER_THREADS=0 \
  -e SGLANG_EXL3_EMBED_HOST=1 \
  -e SGLANG_EXL3_EXPERT_CACHE_GB=auto \
  -e SGLANG_EXL3_HC_INT8=0 \
  -e SGLANG_EXL3_KV_BITS=5 \
  -e SGLANG_EXL3_MM_FAST_CPU=1 \
  -e SGLANG_EXL3_MODEL_PATH=/models/Qwen3.8-Flash-Next-exl3-69e33439 \
  -e SGLANG_EXL3_MOE_OFFLOAD=gpu_cache \
  -e SGLANG_EXL3_MOE_PREFILL_FP16_ACC=1 \
  -e SGLANG_EXL3_NGRAM_RAM_GB=2 \
  -e SGLANG_EXL3_NGRAM_TIER=nvme \
  -e SGLANG_EXL3_OFFLOAD_COMPACT_PARTS=1 \
  -e SGLANG_EXL3_OFFLOAD_CPU_TIER=1 \
  -e SGLANG_EXL3_OFFLOAD_ELASTIC_GB=0 \
  -e SGLANG_EXL3_OFFLOAD_FUSED=1 \
  -e SGLANG_EXL3_OFFLOAD_PREFILL_SUBCHUNK=1073741824 \
  -e SGLANG_EXL3_OFFLOAD_STAGING_PARTS=4 \
  -e SGLANG_EXL3_VISION=0 \
  -e SGLANG_EXL3_VIT_MLP_CHUNK=8192 \
  -e SGLANG_EXL3_VIT_SDPA=1 \
  -v ~/models/Qwen3.8-Flash-Next-exl3-69e33439:/models/Qwen3.8-Flash-Next-exl3-69e33439:ro \
  --entrypoint /opt/entrypoint.sh \
  ghcr.io/sybil-solutions/sglang-exl3-flashnext-multiarch@sha256:fc147f4dd42b89d76406eab482cb3f925346c41e74d983e2ff1dcb721f6d4a2f \
  python3 \
  -m sglang.launch_server \
  --model-path /models/Qwen3.8-Flash-Next-exl3-69e33439 \
  --quantization exl3 \
  --trust-remote-code \
  --host 0.0.0.0 \
  --port 30000 \
  --served-model-name Qwen3.8-Flash-Next-exl3-69e33439 \
  --disable-shared-experts-fusion \
  --context-length 262144 \
  --mem-fraction-static 0.88 \
  --kv-cache-dtype fp8_e4m3 \
  --max-running-requests 1 \
  --chunked-prefill-size 8192 \
  --cuda-graph-max-bs-decode 1 \
  --cuda-graph-backend-prefill disabled \
  --max-mamba-cache-size 5 \
  --max-total-tokens 263168 \
  '--json-model-override-args={}' \
  --reasoning-parser qwen3 \
  --tool-call-parser qwen3_coder

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 5070 Ti on Oct 16, 2026.

Details

Weights: turboderp/Qwen3.8-Flash-Next-exl3 @ 69e33439ae ›
Image: ghcr.io/sybil-solutions/sglang-exl3-flashnext-multiarch@sha256:fc147f4dd42b89d76406eab482cb3f925346c41e74d983e2ff1dcb721f6d4a2f
Engine profile: sglang-exl3-offload-multiarch ›
Recipe file: registry/recipes/nvidia/rtx-5070-ti-16gb/qwen3.8-flash-next.sglang.256k.json ›
Model released: Aug 24, 2026
Tested: vast, RTX 5070 Ti