Qwen3.8-Flash-Next

EXL3 2.05 bpw · tabbyapi · 256K context · RTX PRO 4500 Blackwell · 32 GB

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

23 tok/sdecode

– tok/sprefill

256Kcontext window

Earlier checkSep 19, 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 65c895314393431c09050b2e04e250836b3a6eb4 \
  --local-dir ~/models/Qwen3.8-Flash-Next-exl3-65c89531

2 Write the server config

cat > qwen3.8-flash-next.tabbyapi.256k.yml <<'EOF'
network:
  host: 0.0.0.0
  port: 5000
  disable_auth: true
  disable_fetch_requests: true
  send_tracebacks: false
  api_servers: ["OAI"]
  sse_ping_interval: 15

logging:
  log_prompt: false
  log_generation_params: false
  log_requests: false
  log_chat_completion_requests: false

model:
  model_dir: /workspace/models
  inline_model_loading: false
  use_dummy_models: false
  model_name: Qwen3.8-Flash-Next-EXL3-205bpw-H4
  backend: exllamav3
  max_seq_len: 262144
  cache_size: 262656
  cache_mode: Q4
  tensor_parallel: false
  gpu_split_auto: true
  autosplit_reserve: [512]
  chunk_size: 2048
  output_chunking: true
  max_batch_size: 2
  vision: true
  reasoning: true
  start_in_reasoning: always
  reasoning_start_token: " thinking"
  reasoning_end_token: "</think>"
  template_vars_default:
    enable_thinking: true
  tool_format: qwen3_coder
  # The 26.24 GB n-gram table is indexed by short n-grams and never needs to
  # be on the card, so it lives in host RAM. 51B of routed-expert weights are
  # then split off the card so what stays resident fits a 24 GB board, and the
  # split is chosen from the measured free memory rather than guessed.
  ngram_ram: true
  cpu_moe_split_experts: 128
  draft_model:
    draft_mode: mtp
EOF

3 Start the server

docker run --rm \
  --gpus all \
  -p 8000:5000 \
  --shm-size 8g \
  -e NVIDIA_VISIBLE_DEVICES=all \
  -v ~/models/Qwen3.8-Flash-Next-exl3-65c89531:/workspace/models:ro \
  -v $PWD/qwen3.8-flash-next.tabbyapi.256k.yml:/app/config.yml:ro \
  --entrypoint /opt/venv/bin/python3 \
  ghcr.io/theroyallab/tabbyapi:cu13@sha256:1c13dd17416660856e9962aeee8498827f0acd604eb36795128a03ddfe03c6ef \
  main.py \
  --config /app/config.yml

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 the older check (loads and chats); a full six-check run is pending.

Details

Weights: turboderp/Qwen3.8-Flash-Next-exl3 @ 65c8953143 ›
Image: ghcr.io/theroyallab/tabbyapi:cu13@sha256:1c13dd17416660856e9962aeee8498827f0acd604eb36795128a03ddfe03c6ef
Launch: tabbyapi-qwen3.8-flash-next-2.05bpw-h4-ng4-256k ›
Recipe file: registry/recipes/nvidia/rtx-pro-4500-blackwell-32gb/qwen3.8-flash-next.tabbyapi.256k.json ›
Model released: Aug 24, 2026
Tested: earlier acceptance