A/B: raise max_tokens 256->512 so POCKET's reasoning + final answer complete (not cut)

A/B: show reasoning again (remove no_think prefill); raise A/B max_tokens 96->256 so thinking completes

A/B: one-sentence example prompts (stable tok/s, still fast)

A/B: prefill empty <think></think> to force direct short answers (no reasoning) -> fast races

A/B: apply /no_think so both models answer directly (no <think> block) -> short, fast races

A/B: shorter-answer example prompts + cap max_tokens 200->96 for snappy races

A/B is now the landing tab (swap order); add prism-ml/Bonsai-27B-gguf to models relation

add Tab2 live A/B vs Bonsai (sequential: Bonsai first, then POCKET); 2nd llama-server for Bonsai Q1_0; README=BONSAI vs POCKET

show live tok/s: backend streams predicted_per_second; UI displays live speed + final tok/s under each reply

proxy via native /completion with manual ChatML prompt (bypass embedded template); surface backend errors

use prebuilt llama.cpp release b10091 (qwen35moe, runtime CPU dispatch) - no compile (no OOM), no -march=native (no SIGILL)

fix: cap compile parallelism (-j2) to avoid OOMKilled (exit137) during llama.cpp build

rebuild backend on upstream llama.cpp master (qwen35moe support) via llama-server; FastAPI proxies; portable CPU build

fix: portable CPU build (no -march=native/AVX512) -> stop SIGILL exit132; graceful load-error status

fix: portable CPU build (no -march=native/AVX512) -> stop SIGILL exit132; graceful load-error status

Upload README.md with huggingface_hub

Upload README.md with huggingface_hub

Upload README.md with huggingface_hub

Upload requirements.txt with huggingface_hub

Upload Dockerfile with huggingface_hub

Upload index.html with huggingface_hub

Upload app.py with huggingface_hub

initial commit