GLM-5.3

EXL3 3 bpw · vllm · 320K context · 4 × RTX PRO 6000 Blackwell in one machine · 384 GB

Z.ai's full GLM-5.3, the large sibling of GLM-5.3-Flash: frontier-class reasoning and agent work for a four-card workstation.

81 tok/sdecode

– tok/sprefill

320Kcontext window

Earlier checkSep 26, 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 davidsyoung/GLM-5.3-EXL3-TR3-3.0bpw \
  --revision ffb7770ea91f088eaf7eeef90ca2ed6b26dc70fe \
  --local-dir ~/models/GLM-5.3-EXL3-TR3-3.0bpw-ffb7770e

2 Write the server config

cat > glm-5.3.vllm.320k.4x.yml <<'EOF'
#!/bin/bash
# NCCL allreduce patch (GLM-5.2 winner) + optional extra sed hooks via WRAP_SED (;; separated sed exprs applied to v16)
set -e
D=/tmp/glm52-candidate; mkdir -p $D
cp /usr/local/bin/serve-glm52-v19.sh /usr/local/bin/serve-glm52-v16.sh /usr/local/bin/glm52-pcie-runtime-env.sh /usr/local/bin/glm52-dcp-prefill-policy.sh /usr/local/bin/glm52-pcie-calibration.py $D/
sed -i "s|exec /usr/local/bin/serve-glm52-v16.sh|exec $D/serve-glm52-v16.sh|" $D/serve-glm52-v19.sh
sed -i "s|/usr/local/bin/glm52-pcie-runtime-env.sh|$D/glm52-pcie-runtime-env.sh|g" $D/serve-glm52-v19.sh $D/serve-glm52-v16.sh
if [ "${PCIE_ALLREDUCE:-nccl}" = nccl ]; then sed -i "s/export VLLM_ENABLE_PCIE_ALLREDUCE=1/export VLLM_ENABLE_PCIE_ALLREDUCE=0/" $D/glm52-pcie-runtime-env.sh; fi
if [ -n "${WRAP_SED:-}" ]; then IFS=';;' read -ra EX <<< "$WRAP_SED"; for e in "${EX[@]}"; do [ -n "$e" ] && sed -i "$e" $D/serve-glm52-v16.sh; done; fi
echo "[glm-wrap] effective pcie env:"; grep -n "PCIE_ALLREDUCE" $D/glm52-pcie-runtime-env.sh | head -5
export GLM52_SERVER=$D/serve-glm52-v19.sh
exec /usr/local/bin/serve-gilded-gnosis.sh "$@"
EOF

3 Start the server

docker run --rm \
  --gpus all \
  -p 8000:8000 \
  --ipc host \
  --shm-size 32g \
  -e ASYNC_SCHEDULING=0 \
  -e B12X_PCIE_DMA=0 \
  -e CUDA_DEVICE_ORDER=PCI_BUS_ID \
  -e DCP=4 \
  -e DCP_BACKEND=ag_rs \
  -e DCP_CKV_GATHER=auto \
  -e DCP_CKV_GATHER_MAX_TOKENS=262144 \
  -e DCP_CKV_PREFETCH_DEPTH=auto \
  -e DCP_INDEXER_SHARDS=auto \
  -e DCP_KV_CACHE_INTERLEAVE_SIZE=1 \
  -e DCP_QUERY_SPLIT=auto \
  -e DCP_TOPK_OWNER_MERGE=auto \
  -e GPU_MEMORY_UTILIZATION=0.98 \
  -e GRAPH=32 \
  -e INSTANTTENSOR_BACKEND=BUFFERED \
  -e INSTANTTENSOR_COPY=0 \
  -e KV_CACHE_DTYPE=nvfp4_ds_mla \
  -e KV_FP8_ROPE=1 \
  -e LMCACHE_MODE=off \
  -e LOAD_FORMAT=instanttensor \
  -e MAX_BATCHED_TOKENS=2048 \
  -e MAX_MODEL_LEN=327680 \
  -e MAX_NUM_SEQS=8 \
  -e MODEL=/model \
  -e MODEL_FAMILY=glm52-exl3 \
  -e MOE_BACKEND=b12x \
  -e MOE_MODE=a16 \
  -e MTP=3 \
  -e MTP_DRAFT_SAMPLE_METHOD=probabilistic \
  -e MTP_MOE_BACKEND=triton \
  -e PCIE_CALIBRATION=off \
  -e PYTORCH_CUDA_ALLOC_CONF=expandable_segments:True \
  -e QUANTIZATION=exl3 \
  -e SERVED_MODEL_NAME=glm-5.3 \
  -e TP=4 \
  -e VLLM_B12X_GLM_NOPE_NVFP4=1 \
  -e VLLM_ENABLE_PCIE_ALLREDUCE=0 \
  -e VLLM_ENGINE_READY_TIMEOUT_S=3600 \
  -e VLLM_EXL3_PREFILL_CAPACITY=2048 \
  -e VLLM_EXL3_TRELLIS_MAX_M=64 \
  -e VLLM_LONG_PREFILL_TOKEN_THRESHOLD=4096 \
  -e VLLM_MAX_LONG_PARTIAL_PREFILLS=1 \
  -e VLLM_MAX_NUM_PARTIAL_PREFILLS=2 \
  -e VLLM_USE_B12X_DCP_A2A=0 \
  -v ~/models/GLM-5.3-EXL3-TR3-3.0bpw-ffb7770e:/model:ro \
  -v $PWD/glm-5.3.vllm.320k.4x.yml:/opt/local-ai/glm53-tuned-entrypoint.sh:ro \
  --entrypoint /bin/bash \
  voipmonitor/vllm:infernal-invocation-vllmc53cc73-b12xc0a44a1-fi1ac6942-cu133-torch213-20260817-r17@sha256:c5e96c5bcc5a073f7ce6b56173d88538de3a416900cff97c88b4bf7967fe1dc0 \
  /opt/local-ai/glm53-tuned-entrypoint.sh \
  --default-chat-template-kwargs '{"reasoning_effort":"none"}' \
  --attention-config '{"indexer_kv_dtype":"fp8"}' \
  --kv-cache-memory-bytes 11811160064 \
  --prefill-schedule-interval 2 \
  --enable-prompt-tokens-details

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: davidsyoung/GLM-5.3-EXL3-TR3-3.0bpw @ ffb7770ea9 ›
Image: voipmonitor/vllm:infernal-invocation-vllmc53cc73-b12xc0a44a1-fi1ac6942-cu133-torch213-20260817-r17@sha256:c5e96c5bcc5a073f7ce6b56173d88538de3a416900cff97c88b4bf7967fe1dc0
Launch: vllm-glm-5.3-exl3-3.0bpw-320k-tp4 ›
Recipe file: registry/recipes/nvidia/rtx-pro-6000-blackwell-96gb/glm-5.3.vllm.320k.4x.json ›
Model released: Aug 25, 2026
Tested: earlier acceptance