GLM-5.2

EXL3 3 bpw · vllm · 256K context · 4 × RTX PRO 6000 Blackwell in one machine · 384 GB

Z.ai's GLM-5.2 flagship, here as an NVFP4 + AQLM hybrid.

188 tok/sdecode

– tok/sprefill

256Kcontext window

Earlier checkearlier

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 brandonmusic/GLM-5.2-EXL3-TR3-3.0bpw \
  --revision 677eaa98b41b9f06abaad03953659c2872c22ce3 \
  --local-dir ~/models/GLM-5.2-EXL3-TR3-3.0bpw-677eaa98

2 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=16384 \
  -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 GPUS=0,1,2,3 \
  -e GPU_MEMORY_UTILIZATION=0.95 \
  -e GRAPH=56 \
  -e KV_FP8_ROPE=1 \
  -e LMCACHE_MODE=off \
  -e MAX_BATCHED_TOKENS=3072 \
  -e MAX_MODEL_LEN=262144 \
  -e MAX_NUM_SEQS=8 \
  -e MODEL=/model \
  -e MODEL_FAMILY=glm52-exl3 \
  -e MTP=6 \
  -e MTP_DRAFT_SAMPLE_METHOD=greedy \
  -e MTP_MOE_BACKEND=triton \
  -e NCCL_IB_DISABLE=1 \
  -e NCCL_P2P_LEVEL=SYS \
  -e NCCL_PROTO=LL,LL128,Simple \
  -e PCIE_CALIBRATION=off \
  -e SERVED_MODEL_NAME=glm-5.2 \
  -e TORCH_CUDA_ARCH_LIST=12.0a \
  -e TP=4 \
  -e VLLM_DISABLED_KERNELS=MarlinFP8ScaledMMLinearKernel \
  -e VLLM_DISABLE_SHARED_EXPERTS_STREAM=0 \
  -e VLLM_ENABLE_PCIE_ALLREDUCE=0 \
  -e VLLM_MEMORY_PROFILER_ESTIMATE_CUDAGRAPHS=1 \
  -e VLLM_USE_B12X_DCP_A2A=0 \
  -v ~/models/GLM-5.2-EXL3-TR3-3.0bpw-677eaa98:/model:ro \
  voipmonitor/vllm:gilded-gnosis-v20-vllmb46c3aa-si35aebc6-fi801d57a-cu132-20260730-r12@sha256:bcd65c13076b00cf7804001ce70085453de958b082897b9674a26d9fab4d436a \
  -lc 'mkdir -p /tmp/glm52-candidate && 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 /tmp/glm52-candidate/ && sed -i "s|exec /usr/local/bin/serve-glm52-v16.sh|exec /tmp/glm52-candidate/serve-glm52-v16.sh|" /tmp/glm52-candidate/serve-glm52-v19.sh && sed -i "s/export VLLM_ENABLE_PCIE_ALLREDUCE=1/export VLLM_ENABLE_PCIE_ALLREDUCE=0/" /tmp/glm52-candidate/glm52-pcie-runtime-env.sh && export GLM52_SERVER=/tmp/glm52-candidate/serve-glm52-v19.sh && exec /usr/local/bin/serve-gilded-gnosis.sh'

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: brandonmusic/GLM-5.2-EXL3-TR3-3.0bpw @ 677eaa98b4 ›
Image: voipmonitor/vllm:gilded-gnosis-v20-vllmb46c3aa-si35aebc6-fi801d57a-cu132-20260730-r12@sha256:bcd65c13076b00cf7804001ce70085453de958b082897b9674a26d9fab4d436a
Launch: vllm-glm-5.2-3.0-bpw-256k ›
Recipe file: registry/recipes/nvidia/rtx-pro-6000-blackwell-96gb/glm-5.2.vllm.256k.4x.json ›
Model released: Jun 16, 2026
Tested: earlier acceptance