nvidia NVFP4 (FP8→NVFP4, 2x4 B300)
source ↗MMLU-Pro
87.88+0.07baseline 87.81
GPQA Diamond
92.68+0.89baseline 91.79
LiveCodeBench V6
53.55+0.6baseline 52.95
AIME 2025
99.53+0.15baseline 99.38
Terminal-Bench 2.1
86.8+1.13baseline 85.67
GLM-5.3
- Organization
- Z.ai / THUDM
- Parameters
- 753B total / 40B active
- Context
- 1M
- License
- glm-5.3 (custom)
- Release
- 2026-08-25
GlmMoeDsaForCausalLM, DSA sparse attn, 79 layers, 8 active experts
GLM-5.3 is Z.ai's 753B-total / 40B-active DSA sparse-attention MoE with a 1M-token context, released under the custom glm-5.3 license. The nvidia NVFP4 checkpoint beats its FP8 baseline on 2x4 B300 (GPQA Diamond 91.79 → 92.68), while OneNexus MXFP4 recovers 99.62% of combined BF16 quality on 4x MI350.
GLM-5.3 is a 753B total / 40B active glm-5.3 (custom) model from Z.ai / THUDM with 1M context. Quantized checkpoints lose 0.15–7.3 points on the 5 measured per-quant benchmarks.
Base-model benchmarks
Terminal-Bench 2.1
88.2Terminal-Bench 3.0
28.3DeepSWE v1.1
66.9NL2Repo
58FrontierSWE
78.1CyberGym
84.5Toolathlon Verified
73HLE w/ Tools
62.5ALE-CLI
28.5
Quant checkpoints
| Checkpoint | Family | Size | Source |
|---|---|---|---|
| zai-org FP8 native ~755.7 GB official checkpoint | FP8 | 755.7 GB | source ↗ |
| unsloth GLM-5.3-GGUF UD-IQ1_S | GGUF | 217 GB | source ↗ |
| unsloth GLM-5.3-GGUF Q2_K_XL | GGUF | 254 GB | source ↗ |
| unsloth GLM-5.3-GGUF Q3_K_XL | GGUF | 343 GB | source ↗ |
| unsloth GLM-5.3-GGUF Q4_K_XL | GGUF | 467 GB | source ↗ |
| unsloth GLM-5.3-GGUF Q5_K_XL | GGUF | 562 GB | source ↗ |
| unsloth GLM-5.3-GGUF Q6_K_XL | GGUF | 684 GB | source ↗ |
| unsloth GLM-5.3-GGUF Q8_0 | GGUF | 801 GB | source ↗ |
| OneNexus GLM-5.3-MXFP4 GiB | MXFP4 | 407.92 GB | source ↗ |
| nvidia GLM-5.3-NVFP4 Size not captured | NVFP4 | — | source ↗ |
| gpustack GLM-5.3-W4A8 W4A8 | GGUF | 372.3 GB | source ↗ |
| autotrust GLM-5.3-GGUF-DGX-Spark SLIM-Q E192 IQ2_XXS SLIM-Q E192 IQ2_XXS | GGUF | 149.7 GB | source ↗ |
Per-quant quality deltas
OneNexus MXFP4 (BF16→MXFP4, 4x MI350)
source ↗GSM8K
97.5−0.15baseline 97.65
MMLU
88.6−0.6baseline 89.2
GPQA
53.03−1.01baseline 54.04
W4A8 (gpustack)
source ↗GPQA Diamond
90.28+0.63baseline 89.65
Perplexity
2.44+0baseline 2.44
SLIM-Q E192 IQ2_XXS (autotrust)
source ↗HumanEval
95.1+0baseline 95.1
GPQA Diamond
83.3−3.6baseline 86.9
C-Eval
84.7−7.3baseline 92
Serving metrics
Validated context
1 M tokens (4x MI350 TP4/EP4 EAGLE-MTP)Hardware: 4x AMD MI350
Setup: OneNexus MXFP4
source ↗
What’s not published
- The unsloth GGUF ladder (217–801 GB) has no published accuracy numbers.
- Per-quant deltas exist for NVFP4, MXFP4, W4A8 and the autotrust IQ2_XXS build; the rest of the GGUF ladder has none.
Sources
FAQ
- What is GLM-5.3's license, context window and parameter count?
- glm-5.3 (custom) license, 1M context window, 753B total / 40B active parameters (Z.ai / THUDM).
- How much quality does quantization cost GLM-5.3?
- Across 13 published deltas, quantized checkpoints differ from the full-precision baseline by 0.71 points on average (range -7.3 to 1.13).
- How many quant checkpoints are cataloged for GLM-5.3?
- 12 checkpoints across FP8, GGUF, MXFP4, NVFP4, sizes from 149.7 GB to 801 GB.
- When was GLM-5.3 released?
- 2026-08-25.
- Where do the GLM-5.3 benchmark numbers come from?
- From 6 official sources — vendor model cards, quantization publisher cards and serving recipe pages.