nvidia NVFP4 (BF16→NVFP4, GB200)
source ↗GPQA Diamond
92.11−0.06baseline 92.17
SciCode
57.69+1.48baseline 56.21
MMMU-Pro
76.3−0.58baseline 76.88
IFBench
60.54−0.76baseline 61.3
Terminal-Bench 2.1
83.15+0.57baseline 82.58
GLM-5.3-Flash
- Organization
- Z.ai / THUDM
- Parameters
- 320B total / 18B active
- Context
- 1M
- License
- MIT
- Release
- 2026-08-25
MoE, 45 layers + 1 MTP, 288 experts + 1 shared, top-8, KDA linear-attn + sparse MLA
GLM-5.3-Flash is Z.ai's 320B-total / 18B-active MoE with KDA linear attention and sparse MLA, shipping a 1M-token context under the MIT license. Its quantization story centers on the nvidia NVFP4 checkpoint, which holds BF16 quality on a GB200 (GPQA Diamond 92.17 → 92.11) while the turboderp EXL3 build at 2.05 bpw (85.2 GB) reaches 41.8 tok/s on two RTX Pro 6000 cards.
GLM-5.3-Flash is a 320B total / 18B active MIT model from Z.ai / THUDM with 1M context. Quantized checkpoints lose 0.06–0.76 points on the 3 measured per-quant benchmarks.
Base-model benchmarks
Terminal-Bench 2.1
84.3DeepSWE v1.1
63.4HLE w/ tools
55.3ALE-CLI
26.3GPQA Diamond (BF16)
92.17SciCode (BF16)
56.21MMMU-Pro (BF16)
76.88IFBench (BF16)
61.3Terminal-Bench 2.1 (BF16)
82.58
Quant checkpoints
| Checkpoint | Family | Size | Source |
|---|---|---|---|
| zai-org FP8 native ~306 GiB official checkpoint | FP8 | 306 GB | source ↗ |
| amd GLM-5.3-Flash-Quark-MXFP4 | MXFP4 | 185 GB | source ↗ |
| OneNexus GLM-5.3-Flash-MXFP4 | MXFP4 | 195.6 GB | source ↗ |
| nvidia GLM-5.3-Flash-NVFP4 Size not captured | NVFP4 | — | source ↗ |
| unsloth GLM-5.3-Flash-GGUF UD-Q2_K_XL ~109 GB | GGUF | 109 GB | source ↗ |
| unsloth GLM-5.3-Flash-GGUF Q3_K_XL | GGUF | 137.4 GB | source ↗ |
| unsloth GLM-5.3-Flash-GGUF Q4_K_XL | GGUF | 185.99 GB | source ↗ |
| ggml-org GLM-5.3-Flash Q4_K | GGUF | 168.92 GB | source ↗ |
| turboderp GLM-5.3-Flash EXL3 2.05bpw | EXL3 | 85.2 GB | source ↗ |
| autotrust GLM5.3-Flash-E256-DGX-Spark NVFP4 NVFP4 E256 expert-pruned | NVFP4 | 170.5 GB | source ↗ |
| autotrust GLM5.3-Flash-E224-DGX-Spark NVFP4 NVFP4 E224 expert-pruned | NVFP4 | 151.5 GB | source ↗ |
| canada-quant GLM-5.3-Flash-W4A16-MTP W4A16-MTP | GGUF | 177.7 GB | source ↗ |
| AxonZeta GLM-5.3-Flash-AZ-Focus-EXL3 EXL3 AZ-Focus; size not captured | EXL3 | — | source ↗ |
Per-quant quality deltas
NVFP4 E256 expert-pruned (autotrust)
source ↗HumanEval
97.6GPQA Diamond
77.3C-Eval
89.4AIME 2025
74.2MMMU
76.1NVFP4 E224 expert-pruned (autotrust)
source ↗GPQA Diamond
90.9AIME 2025
88.3HumanEval
98.2C-Eval
84MMMU
73.6W4A16-MTP (canada-quant)
source ↗AIME 2025
88.33GSM8K
97GPQA Diamond
85.86MMMU
74.67EXL3 AZ-Focus (AxonZeta)
source ↗MMLU-Pro
86.8+0.6baseline 86.2
AIME 2025
81.7+6.7baseline 75
GPQA Diamond
82.3+1.5baseline 80.8
HumanEval+
86.9+1.1baseline 85.8
Serving metrics
KV pool @128K (FP8 vs BF16)
14.92 M tokens (vs 8.87M BF16)Hardware: AMD MI355X
Setup: vLLM FP8 TP4
source ↗KV capacity
1.8 x (SGLang FP8 KV, +2.9–5.7% throughput)Hardware: NVIDIA GB300
Setup: SGLang FP8 KV cache
source ↗EXL3 decode
41.8 tok/sHardware: 2x NVIDIA RTX Pro 6000
Setup: turboderp EXL3 2.05bpw (TabbyAPI)
source ↗FP8 single-tray fit
1 GB200 tray (vLLM FP8 TP4 + MTP5)Hardware: NVIDIA GB200
Setup: vLLM FP8 TP4 + MTP5
source ↗
What’s not published
- The official full benchmark table is image-only on the model card.
- No per-quant quality table is published for the MXFP4 and GGUF checkpoints — the autotrust, canada-quant and AxonZeta cards publish single-score or paired tables only.
Sources
- https://huggingface.co/zai-org/GLM-5.3-Flash-BF16 ↗
- https://huggingface.co/amd/GLM-5.3-Flash-Quark-MXFP4 ↗
- https://huggingface.co/OneNexus/GLM-5.3-Flash-MXFP4 ↗
- https://huggingface.co/nvidia/GLM-5.3-Flash-NVFP4 ↗
- https://huggingface.co/unsloth/GLM-5.3-Flash-GGUF ↗
- https://huggingface.co/ggml-org/GLM-5.3-Flash-GGUF ↗
- https://huggingface.co/turboderp/GLM-5.3-Flash-exl3 ↗
- https://huggingface.co/autotrust/GLM5.3-Flash-E256-DGX-Spark ↗
- https://huggingface.co/autotrust/GLM5.3-Flash-E224-DGX-Spark ↗
- https://huggingface.co/canada-quant/GLM-5.3-Flash-W4A16-MTP ↗
- https://huggingface.co/AxonZeta/GLM-5.3-Flash-AZ-Focus-EXL3 ↗
- https://recipes.vllm.ai/zai-org/GLM-5.3-Flash ↗
FAQ
- What is GLM-5.3-Flash's license, context window and parameter count?
- MIT license, 1M context window, 320B total / 18B active parameters (Z.ai / THUDM).
- How much quality does quantization cost GLM-5.3-Flash?
- Across 9 published deltas, quantized checkpoints differ from the full-precision baseline by 1.17 points on average (range -0.76 to 6.7).
- How many quant checkpoints are cataloged for GLM-5.3-Flash?
- 13 checkpoints across FP8, MXFP4, NVFP4, GGUF, EXL3, sizes from 85.2 GB to 306 GB.
- When was GLM-5.3-Flash released?
- 2026-08-25.
- Where do the GLM-5.3-Flash benchmark numbers come from?
- From 12 official sources — vendor model cards, quantization publisher cards and serving recipe pages.