LLMCheck: Apple M3 Ultra (512GB), LM Studio, GLM 5.2 753B-A40B IQ1_S decode
22.0 tok/s (decode, per-request)
Fields
id
benchmarks/llmcheck-m3-ultra-glm5-753b-iq1s-lmstudio-decodeMetric
decode_tok_s
Value
22
Unit
tok/s (decode, per-request)
Model
GLM 5.2 753B-A40B
Methodology
THIRD-PARTY MEASURED (LLMCheck open dataset). M3 Ultra 512GB, LM Studio, GLM 5.2 753B-A40B at IQ1_S (extreme ~1.7-bit quant), TTFT 1.5 s, per-request decode.
Measured by
third_party
Reproducible
yes
notes
Extreme-quant data point: a 753B MoE only fits at IQ1_S on 512GB, and still decodes at reading speed (22 tok/s). The Q4_K_M MLX row for the same model is 15 tok/s — recorded separately. Note quant IQ1_S has no dedicated quantization record in the repo.
Notes
Extreme-quant data point: a 753B MoE only fits at IQ1_S on 512GB, and still decodes at reading speed (22 tok/s). The Q4_K_M MLX row for the same model is 15 tok/s — recorded separately. Note quant IQ1_S has no dedicated quantization record in the repo.