snapshot 2026.06: vision ladder, 2 devices, fp32/fp16/int8

README.md

@@ -40,6 +40,17 @@ capacity indicator; rankings and deltas are the signal.
 - **Published accuracy can invert on-device rankings:** EfficientNet-B2
   (published +2.9 over B0) measured *no better* than B0 while 51% slower.
 
+## LLM benchmarks (llm-benchmarks.csv)
+
+GGUF quantization ladders measured on-device via llama.cpp (CPU): decode and
+prefill tok/s at the measured-optimal thread count (big cores only — the
+llama.cpp default of all-cores is 28-39% slower on big.LITTLE SoCs), with
+WikiText-2 perplexity (fixed 32-chunk subset) as the device-independent
+quality column. First snapshot: Qwen2.5-0.5B-Instruct, Q8_0→Q3_K_M, on a
+Snapdragon 662 phone — 14-19 tok/s decode (conversational) at every level;
+Q6_K dominates Q8_0; prefill (~30-40 tok/s) is the practical bottleneck for
+long prompts.
+
 ## Schema
 
 One row = one (model, device, precision/quant) measurement.

llm-benchmarks.csv

@@ -0,0 +1,6 @@
+snapshot,task,model,params_b,device,device_type,soc,quant,file_mb,runtime,threads,decode_tok_s,prefill_tok_s,quality_metric,quality_value,sweep
+2026.06,text-generation,qwen2.5-0.5b-instruct,0.5,oppo-a74,phone,Snapdragon 662,q8_0,675.7,"llama.cpp b9590 (Termux, CPU)",4,13.91,38.51,wikitext2_ppl_32chunks,13.8042,llm-quant-v1
+2026.06,text-generation,qwen2.5-0.5b-instruct,0.5,oppo-a74,phone,Snapdragon 662,q6_k,650.4,"llama.cpp b9590 (Termux, CPU)",4,15.59,35.97,wikitext2_ppl_32chunks,13.8091,llm-quant-v1
+2026.06,text-generation,qwen2.5-0.5b-instruct,0.5,oppo-a74,phone,Snapdragon 662,q5_k_m,522.2,"llama.cpp b9590 (Termux, CPU)",4,16.58,29.74,wikitext2_ppl_32chunks,13.9929,llm-quant-v1
+2026.06,text-generation,qwen2.5-0.5b-instruct,0.5,oppo-a74,phone,Snapdragon 662,q4_k_m,491.4,"llama.cpp b9590 (Termux, CPU)",4,17.69,33.41,wikitext2_ppl_32chunks,14.3045,llm-quant-v1
+2026.06,text-generation,qwen2.5-0.5b-instruct,0.5,oppo-a74,phone,Snapdragon 662,q3_k_m,432.0,"llama.cpp b9590 (Termux, CPU)",4,19.25,35.73,wikitext2_ppl_32chunks,14.7384,llm-quant-v1