GLM-5.3 · NVFP4

753B total · 256 routed experts · DeepSeek-sparse attention · 1M context

1403 GiB → 433 GiB  ·  −69%  ·  round-trip cosine 0.9961

NVFP4 quantization of Z.ai's GLM-5.3, following NVIDIA's own GLM-5.2 recipe tensor-for-tensor.

Quantized by LibertAI · not affiliated with Z.ai / Zhipu / NVIDIA

✨ What this is

An NVFP4 checkpoint of GLM-5.3. The routed-expert FFN tensors — 57,600 of them, the overwhelming majority of the model's weight — are quantized to NVFP4 (E2M1, FP8-E4M3 per-16-block scales, FP32 per-tensor global scale). Everything outlier-sensitive stays in BF16:

all DeepSeek-sparse attention including the sparse indexer · all shared experts · the MoE routers · the three dense layers (0–2) · the entire MTP layer 78, its experts included · token embeddings · lm_head · all norms

Activations are not calibrated — see the input_scale note below, which is the one genuinely interesting thing we found while making this.

🎯 The recipe is not ours — and that is the point

Z.ai state it plainly in the GLM-5.3 model card:

"GLM-5.3 uses the same base model as GLM-5.2 — every gain comes from post-training."

The checkpoints bear that out exactly: GLM-5.3 and GLM-5.2 are both glm_moe_dsa, both 753,329,940,480 parameters, with identical config.json geometry. So rather than invent a partition, we reproduce the one NVIDIA ships in nvidia/GLM-5.2-NVFP4 — a checkpoint that has been serving production traffic on our 4×B200 box since June 2026.

We verified the reproduction two ways before quantizing a single weight:

And once the weights were built, the finished checkpoint closed the loop: its summed total_size is 464,795,267,072 bytes — byte-identical to NVIDIA's. Same names, same shapes, same dtypes, same byte count.

Concretely, what gets quantized is:

model.layers.{3..77}.mlp.experts.{0..255}.{gate,up,down}_proj     # 75 × 256 × 3 = 57,600

…and nothing else. Layers 0–2 are dense (first_k_dense_replace: 3) and layer 78 is the MTP head; NVIDIA keeps both entirely in BF16, so we do too.

🔬 About input_scale

Each quantized tensor here ships four entries — .weight, .weight_scale, .weight_scale_2, and .input_scale. That last one is exactly 1.0 for all 57,600 tensors, and that is deliberate.

We sampled 24 .input_scale values out of nvidia/GLM-5.2-NVFP4 by HTTP range-read. Every one is 1.0. NVIDIA — who wrote ModelOpt — performed no activation calibration for that checkpoint; the tensors are unit placeholders. So in substance this is a weight-only NVFP4 quantization, and we did not fabricate a calibration we did not run.

But the tensors must still be present. vLLM's ModelOptNvFp4FusedMoE folds input_scale into the dequantization alpha:

alpha = weight_scale_2 * input_scale

When the tensor is absent, vLLM folds an uninitialised PerTensorScaleParameter — observed as 0.0 — which zeroes every expert and makes the model emit garbage. We hit exactly this on our own earlier weight-only GLM-5.3-Flash checkpoint and filed it as vllm#54189. Emitting 1.0 makes the fold a no-op and is precisely what the production-proven NVIDIA checkpoint does.

If you are quantizing an NVFP4 MoE yourself: emit input_scale, even if you did not calibrate. That is the takeaway.

🖥 Hardware

The FP8 release (~700 GiB) needs 8× B200; this NVFP4 build is what makes GLM-5.3 a 4-GPU model.

🚀 Serving

Unlike GLM-5.3-Flash — a genuinely new glm5_next architecture that needed per-model engine images — GLM-5.3 is glm_moe_dsa, which vLLM and SGLang already support. No custom build, no patches.

vllm serve LibertAIDAI/GLM-5.3-NVFP4 \
  --tensor-parallel-size 4 \
  --max-model-len 262144 \
  --kv-cache-dtype fp8 \
  --enable-auto-tool-choice --tool-call-parser glm47 \
  --reasoning-parser qwen3 \
  --served-model-name glm-5.3

Notes carried over from running its GLM-5.2 sibling in production:

• --tool-call-parser glm47 and --reasoning-parser qwen3 are the right pair for this family.
• 64 attention heads divide cleanly by TP 2 / 4 / 8.
• reasoning_effort accepts low, high, max — and defaults to max, so budget max_tokens generously.
• The chat template's clear_thinking defaults to false; pass clear_thinking=true for chat-style use.

🔧 How it was made

CPU-only shard-streaming pass with NVIDIA ModelOpt — each of the 282 source shards is fetched, quantized, written, and dropped, so the job never holds more than a few shards at once and needs nowhere near 1.37 TiB of scratch.

Source is zai-org/GLM-5.3-BF16, not the FP8 release: the FP8 repo carries a modules_to_not_convert list, the signature of a converted checkpoint, so BF16 is the original and quantizing from it avoids stacking two lossy conversions.

Round-trip fidelity, measured by dequantizing back to BF16 and taking cosine similarity against the source weights: mean 0.9961 across sampled expert tensors.

The quantizer, the config builder, and the validation harness are published in full: server-admin/servers/glm53-nvfp4-files.

🙏 Credits

Z.ai for GLM-5.3 and for releasing a BF16 checkpoint alongside the FP8 · NVIDIA for ModelOpt and for the GLM-5.2 recipe this follows · quantized and published by LibertAI.

Quantization only — no weights were retrained, distilled, or otherwise modified beyond the numeric format described above.