QuantThink Leaderboard

Real GPU results measuring how weight- and KV-cache quantization degrade small reasoning (long chain-of-thought) models on a 4GB consumer GPU — and, given a fixed VRAM budget, which config maximizes accuracy.

GitHub repo Results dataset Suite dataset GGUF model GPTQ model
Memory-Budget Frontier: under the accuracy objective, the optimal model family crosses over as the VRAM budget grows — DeepSeek-R1-Distill-Qwen-1.5B wins below ~2.3GB, Qwen3-1.7B wins above it. Under the Cost-to-Solve objective, the smaller model wins the entire 2–4GB grid instead, because the larger model's accuracy gain doesn't pay for its higher token cost. Separately, Q4 KV-cache quantization caused total generation collapse (not smooth degradation) for the R1-distill model at Q4_K_M weights — confirmed by inspecting raw output. First-pass, disclosed-small-N result (N=4–12 samples per cell); see docs/RUN_REAL.md for the full write-up.

GSM8K, real GPU results

Model Quant Acc TL CTS VRAM
DeepSeek-R1-Distill-Qwen-1.5B fp16 0.667 469.8 1541.5 3.50 GB
DeepSeek-R1-Distill-Qwen-1.5B Q8_0 0.750 385.5 1246.4 2.16 GB
DeepSeek-R1-Distill-Qwen-1.5B Q5_K_M 0.667 244.3 985.5 1.67 GB
DeepSeek-R1-Distill-Qwen-1.5B Q4_K_M 0.583 420.8 1532.1 1.55 GB
DeepSeek-R1-Distill-Qwen-1.5B GPTQ 4-bit (self-calibrated) 0.750 317.4 1025.1 1.63 GB*
Qwen3-1.7B (thinking) Q8_0 1.000 1191.8 2212.9 2.99 GB
Qwen3-1.7B (thinking) Q4_K_M 0.875 1910.1 3717.6 2.32 GB
Qwen3-0.6B (thinking) fp16 0.500 1300.0 4934.3 2.40 GB
Qwen3-0.6B (thinking) Q4_K_M 0.375 1385.5 7455.3 1.65 GB

* GPTQ VRAM measured via torch.cuda.max_memory_allocated, not directly comparable to the nvidia-smi-based GGUF measurements. Full table (including MATH-500, KV-cache axis, thinking-cap grid) in docs/RUN_REAL.md. Acc = pass@1. TL = mean thinking-length (tokens). CTS = cost-to-solve (expected tokens per correct answer).

Related projects