results.csv

model,format,file_bytes,bpw_all,wikitext2_ppl,ratio_to_fp16,zero_shot_avg,note,source
Qwen3-4B,fp16 (llama.cpp),8051285408,16.013,12.4567,1.000,,"reference; WikiText-2 test, ctx 4096, 73 windows",paper: main equal-bytes table
Qwen3-4B,Q8_0 (llama.cpp),4280405408,8.513,12.4978,1.003,56.27,"8-bit reference; stands in for fp16 in the zero-shot runs; WikiText-2 test, ctx 4096, 73 windows",paper: stock-quant curve table
Qwen3-4B,"Pingala-8 (B4, ours)",1243153792,2.472,16.6413,1.336,50.83,"ours; post hoc (last of a chain of builds evaluated on WikiText-2); needs our patched llama.cpp (CUDA); WikiText-2 test, ctx 4096, 73 windows",paper: main equal-bytes table
Qwen3-4B,"IQ2_KT, C4-calibrated (ik_llama.cpp)",1245914752,2.478,18.1480,1.457,46.52,"best of the six IQ2_KT files we built at this size; WikiText-2 test, ctx 4096, 73 windows",paper: main equal-bytes table
Qwen3-4B,"IQ2_KT, tuned (ik_llama.cpp)",1245675136,2.477,18.2721,1.467,,"best of three recipes tuned for perplexity; WikiText-2 test, ctx 4096, 73 windows",paper: main equal-bytes table
Qwen3-4B,"IQ2_KT, default recipe (ik_llama.cpp)",1300786816,2.587,17.2750,1.387,,"4.6% larger file than ours; WikiText-2 test, ctx 4096, 73 windows",paper: main equal-bytes table
Qwen3-4B,IQ2_XXS (llama.cpp),1246621312,2.479,28.6083,2.297,41.86,"stock llama.cpp, same fp16 file and importance matrix; 0.3% larger file than ours; WikiText-2 test, ctx 4096, 73 windows",paper: main equal-bytes table
Qwen3-4B,IQ2_XS (llama.cpp),1354100352,2.693,19.0193,1.527,47.74,"stock llama.cpp; 8.9% larger file than ours; WikiText-2 test, ctx 4096, 73 windows",paper: stock-quant curve table
Qwen3-4B,Q2_K (llama.cpp),1669500032,3.320,15.0133,1.205,,"stock llama.cpp; 34.3% larger file than ours; WikiText-2 test, ctx 4096, 73 windows",paper: stock-quant curve table
Qwen3-4B,IQ1_M (llama.cpp),1127018112,2.241,57.4988,4.616,,"stock llama.cpp; 9.3% smaller file than ours; WikiText-2 test, ctx 4096, 73 windows",paper: stock-quant curve table
Qwen3-4B,GSQ Q2_K (ISTA-DASLab GGUF),1669499744,3.320,14.3723,1.154,,"third-party compression of Qwen3-4B, run in our harness; 34.3% larger file than ours; WikiText-2 test, ctx 4096, 73 windows",paper: same-protocol comparison table
Qwen3-4B,Ternary Bonsai 4B (Prism ML),1069018824,2.126,11.5883,0.930,,"retrained ternary model, not a compression of the original Qwen3-4B weights; perplexity measured on a Q8_0 copy of its 16-bit release; bytes calculated from tensor shapes for its Q2_0 file, which needs the authors' llama.cpp fork; WikiText-2 test, ctx 4096, 73 windows",paper: same-protocol comparison table
Llama-3.1-8B,"Pingala-8 (B4, ours)",2388210816,,8.44,,,"pre-registered; WikiText-2 test, ctx 4096, 70 windows, BOS at every window",paper: Llama-3.1-8B section (BOS at every window)
Llama-3.1-8B,IQ2_KT (ik_llama.cpp),2389350432,,9.36,,,"quantized from Q8_0; WikiText-2 test, ctx 4096, 70 windows, BOS at every window",paper: Llama-3.1-8B section (BOS at every window)
Llama-3.1-8B,"Pingala-8 (B4, ours)",2388210816,,9.14,,,"pre-registered; WikiText-2 test, ctx 4096, 70 windows, llama.cpp strided mode (windows after the first start without a BOS)",paper: main equal-bytes table
Llama-3.1-8B,IQ2_KT (ik_llama.cpp),2389350432,,10.04,,,"quantized from Q8_0; WikiText-2 test, ctx 4096, 70 windows, strided mode (no BOS after the first window)",paper: main equal-bytes table
Llama-3.1-8B,IQ2_XXS (llama.cpp),2399212576,,12.35,,,"WikiText-2 test, ctx 4096, 70 windows, strided mode (no BOS after the first window)",paper: main equal-bytes table
Qwen3-14B,"Pingala-8 (B4, ours)",4288702112,,9.5684,,,"pre-registered; LOSES to IQ2_KT on WikiText-2 (tie on C4); made from Q8_0; WikiText-2 test, ctx 4096, 73 windows",paper: main equal-bytes table
Qwen3-14B,IQ2_KT (ik_llama.cpp),4288068800,,9.3985,,,"made from Q8_0; WikiText-2 test, ctx 4096, 73 windows",paper: main equal-bytes table
Qwen3-14B,IQ2_XXS (llama.cpp),4299192000,,10.5413,,,"made from Q8_0; WikiText-2 test, ctx 4096, 73 windows",paper: main equal-bytes table