GLM-5.3-Flash, quantized for two RTX PRO 6000 Blackwell GPUs

A mixed-precision quantization of zai-org/GLM-5.3-Flash-BF16 (revision f12e0fe1f6b2ea274c11a569582edfd99d993c5e) that fits two 96 GB SM120 GPUs with a 499,712-token context, four concurrent requests, speculative decoding and vision input. It scores 96.9% on GSM8K (1,278 of 1,319, zero-shot, greedy), within repetition spread of the BF16 source served the same way.

It is served by the matching ormandj/sglang-glm53-flash-sm120 image. It is not an upstream Z.ai, ModelOpt or SGLang release.

Run it

You need Linux x86_64, two visible SM120 GPUs, Docker with the NVIDIA Container Toolkit and about 170 GB of disk.

pip install -U huggingface_hub
export MODEL_DIR=/srv/models/GLM-5.3-Flash-W4A16-NVFP4-K32-Experts-FP8-WO
HF_XET_HIGH_PERFORMANCE=1 hf download ormandj/GLM-5.3-Flash-W4A16-NVFP4-K32-Experts-FP8-WO --local-dir "$MODEL_DIR"

git clone https://github.com/ormandj/sglang-glm53-flash-sm120
cd sglang-glm53-flash-sm120
export IMAGE=ghcr.io/ormandj/sglang-glm53-flash-sm120:v0.1.1
export CACHE_DIR=/srv/cache/sglang-glm53-flash-sm120-v55
./examples/serve-glm53-flash.sh

The first boot compiles kernels into CACHE_DIR and takes about 10 to 20 minutes; later boots take about 8. The server is OpenAI-compatible on port 8000 with model name glm-5.3-flash. Measured throughput, capacity and quality are in the repository's README.

What is quantized

• Routed experts in layers 3 through 45: ModelOpt W4A16 NVFP4, signed E2M1 packed weights, E4M3FN scales over K=32, BF16 activations.
• 317 attention and shared-expert projections in layers 0 through 44: E4M3 block-FP8 weight-only with FP32 inverse scales over 128x128 blocks.
• Kept in source precision: the layer 45 MTP draft weights (quantizing them silently collapses speculative acceptance to zero), the DSA indexer, the vision tower and projector, embeddings, LM head, routers, norms and ineligible linears.
• Tensor payload: 177,995,252,856 bytes in 90 safetensors shards.

Reconstruction against the BF16 source: routed experts aggregate relative L2 0.0851 (minimum matrix cosine 0.9949); FP8 tier aggregate relative L2 0.0225 (minimum cosine 0.9996). The full tensor selection, producer pins, per-tensor measurements and per-shard hashes are in quantization-manifest.json, reconstruction-metrics-*.json and model-sha256.json.

Runtime compatibility

The checkpoint uses the modelopt_mixed metadata contract and needs the pinned SGLang and FlashInfer integration in the image above (--quantization modelopt_mixed, shared-expert fusion disabled). Upstream Transformers, vLLM and SGLang releases do not load it correctly today.

Integrity

The publishing build refuses to upload unless its model files reproduce the served artifact exactly:

• config.json: 00cf7f0ae5cbc1ca1593376730aa52dd918b59cab28552c61e9d8e3fd7a19d20
• model.safetensors.index.json: 944bb41dd3dc2b6fde6b37465b9d9f7743ea89002302b7c38a87e0242270554d
• model-sha256.json: 655a9b802f0629d1067f2350ee4e6d4c288c642846f2b1955760000ebf9bdfa5

The last file holds the SHA-256 digest of every model shard.

License and attribution

The source checkpoint is MIT-licensed. See the included license and the upstream model card for the original model's usage guidance, limitations and citation.