4× RTX PRO 6000
81.7tok/sEveryday
1 request, prose prompt, with MTP Source
Intelligence34with thinkingno thinking 22
- ≤ 3 bit: quantization may cost quality
Local AI · Recipe · 4× RTX PRO 6000
GLM-5.2 EXL3 3.0 bpw with vLLM on 4× RTX PRO 6000 (500k context, text)
GLM-5.2 EXL3 3.0 bpw with vLLM on 4× RTX PRO 6000 (500k context, text): 104 tok/s according to github.com (dataset as of Nov 3, 2026).
by jcartu
Intelligence (original model) · Artificial Analysis34with thinkingno thinking 22
≤ 3 bit: quantization may cost quality
Engine
- Engine
- vLLM
- Quantization
- EXL3/Trellis 3.0 bpw, nvfp4_ds_mla KV
- Model family
- GLM-5.2
- Context
- 512,000
- Creator
- jcartu
- GitHub stars
- 0
- Repo updated
- Jul 28, 2026
Measurements
Every sourced value of this recipe, each with its condition and source. Bars relative to the largest value in the group.
Decode · 1 request
Prefill by context
What you need
- Hardware
- 4 × NVIDIA RTX PRO 6000 Blackwell
- Weights
- Main weightsbrandonmusic/GLM-5.2-EXL3-TR3-3.0bpw MIT
- Engine
- vLLM
- Context
- 512,000 tokens
Notes
What matters before you rebuild it.
- ≤ 3 bit: quantization may cost qualityAt 3 bit and below the model may answer noticeably worse than the original. The intelligence number refers to the original.
Sources
- github.com/jcartu/glm52-exl3-blackwell-inference-study/blob/main/ISSUE33_STUDY.md#full-sustained-decode-comparison
- github.com/jcartu/glm52-exl3-blackwell-inference-study/blob/main/ISSUE33_STUDY.md#final-production-versus-matched-stock
- github.com/jcartu/glm52-exl3-blackwell-inference-study
- github.com/jcartu/glm52-exl3-blackwell-inference-study/blob/main/ISSUE33_STUDY.md
- github.com/jcartu/glm52-exl3-blackwell-inference-study/blob/main/results/issue33-upgrade/production-final-dcp2-g9675-owner0-chunk1-pf64-sh16-ckv140k-b3072-max512k-inference-temp1.json
- github.com/jcartu/glm52-exl3-blackwell-inference-study/blob/main/configs/docker-compose.exl3-v20.yml
Related recipes
4× RTX PRO 6000
70.7tok/sEveryday
1 request, prose prompt, with MTP Source
Intelligence34with thinkingno thinking 22
- ≤ 3 bit: quantization may cost quality
4× RTX PRO 6000
64tok/sEveryday
1 request, prose prompt, with MTP (per recipe) Source
Intelligence34with thinkingno thinking 22
- ≤ 3 bit: quantization may cost quality
- Custom kernel required
4× RTX PRO 6000
58.5tok/sEveryday
1 request, prose prompt, with MTP (per recipe) Source
Intelligenceno independent value
- Custom kernel required
4× RTX PRO 6000
92.5tok/sMixed
1 request, mixed prompt set, with MTP Source
Intelligenceno independent value
4× DGX Spark
23.7tok/sMixed
1 request, mixed prompt set, with MTP (per recipe) Source
Intelligence34with thinkingno thinking 22
- Experimental
- Custom kernel required
Local AI in your company?
In a workshop we work out which models and which hardware fit your tasks, and build the first agent on your infrastructure.