BBW alpha — interactive pareto
45 arms · 86 tasks · pi 0.85.1 · frontier is drawn
within each VRAM envelope
UnbiasedStrata R2 passed the full task, engine and log audits and participates in the frontier.
Its serving engine, INT8 KV, 192k context and uncapped reasoning differ from the llama-server runs;
see the config and R2 task audit .
The earlier R1 remains provisional because of serving and harness defects and does not set the frontier.
R1 audit .
UnbiasedStrata Flash-Next IQ3_XXS +MTP UnbiasedStrata 2 Flash-Next IQ3_S +MTP (128k, q4_0 KV) UnbiasedStrata Flash-Next IQ3_XXS +MTP (16 GiB VRAM, arena, 128k) Muse Spark Open 816B-A42B IQ3_XXS (2× B200) Qwen3.6-27B IQ4_XS UnbiasedStrata Flash-Next IQ3_XXS +MTP (rotated Q4 KV) UnbiasedStrata Coder IQ1_M +MTP (32 GiB RAM) UnbiasedStrata Flash-Next IQ3_XXS (provisional) UnbiasedStrata Flash-Next IQ2_XS +MTP (16 GiB VRAM) ThinkingCap-3.6-27B Q4_K_M +MTP UnbiasedStrata Flash-Next IQ3_XXS +MTP (16 GiB VRAM, 64 GiB RAM, arena) UnbiasedStrata Coder IQ1_M +MTP (16 GiB VRAM, 3090) Muse-Glimmer-30B IQ3_XXS Swift 1.5 Qwen3.8-27B GSQ-RCO IQ3_S +MTP Qwen3.6-27B UD-IQ3_XXS +MTP (16 GiB, pod 4090) ThinkingCap-3.8-27B Q4_K_M +MTP (medium) Qwen3.8-27B GSQ-RCO IQ3_S +MTP UnbiasedStrata Flash-Next Q2_0 +MTP (16 GiB VRAM, 32 GiB RAM) Qwen3.6-35B-A3B IQ4_XS Qwen3.6-35B-A3B Q4_K_M +MTP 256k expert-LRU Qwen3.8-27B Q3_K_M Swift 1.5 Flash-Next Q2_0 +MTP (16 GiB VRAM, 32 GiB RAM) Qwen3.6-35B-A3B IQ3_XXS (8GB) Saluki-27B GSQ-RCO IQ2-mix Qwen3.6-35B-A3B REAP-48 (8GB) Gemma-4-26B-A4B Q4_K_M Qwen3.6-35B-A3B IQ2_M (8GB) Bonsai-2-27B PTQ1_0 256k pen0 Heretic 35B-A3B Q3_K_M 256k Bonsai-2-27B PTQ1_0 256k Bonsai-2-27B PTQ1_0 48k (8GB) Qwen3.5-9B Q4_K_M (8GB) HauhauCS 35B-A3B IQ2_M GLM-4.7-Flash Q4_K_M 256k Qwen3.5-9B Q4_K_M Gemma-4-26B-A4B IQ3_XXS (8GB) HauhauCS 35B-A3B IQ2_M (8GB) LFM3-1.7B Q8_0 f16 KV 256k (8GB) Xing4.0-29B-A4B IQ4_NL +MTP Gemma-4-12B IQ4_XS 40k (8GB) Bonsai-27B Q1_0 128k (8GB) Granite-4.2-8B Q4_K_M 48k (8GB) Qwen3.8-27B IQ2_XXS 16k (8GB) Huihui 35B-A3B Q2_K Qwen3.8-27B UD-IQ1_S 256k wall-clock hours to work the suite (lower is better)
mean score (higher is better)
No arm selected
Click any point to pin its configuration here. Shift-click to
add it to a comparison. Hover to preview without pinning.
clear focus
view:
−
reset zoom
+
fit visible
1.0×
scroll to zoom · drag to pan · double-click zooms in · +/−/0 keys
find:
engineall llama.cpp sugmeoff unbiasedstrata unbiasedstrata2 quantall GSQ-RCO-Coder-IQ1_M GSQ-RCO-IQ2-mix GSQ-RCO-IQ2_XS GSQ-RCO-IQ3_S GSQ-RCO-IQ3_XXS GSQ-RCO-Q2_0 IQ2_M IQ3_S IQ3_XXS IQ4_NL IQ4_XS PTQ1_0 Q1_0 Q2_K Q3_K_M Q4_K_M Q8_0 REAP-48-v2 Swift-1.5-GSQ-RCO-Q2_0 UD-IQ1_S UD-IQ2_M UD-IQ2_XXS UD-IQ3_XXS UD-IQ4_XS UD-Q4_K_M familyall bonsai-2 bonsai-qwen3.6 gemma glm granite gsqrco hauhau-qwen3.6 heretic-qwen3.6 huihui-qwen3.6 lfm3 muse-glimmer muse-spark-open qwen3.5 qwen3.6 qwen3.6-reap qwen3.8 qwen3.8-flash-next saluki swift15 thinkingcap xing4 contextall 128k 16k 192k 256k 40k 48k 96k GPUall 2x NVIDIA B200 NVIDIA GeForce RTX 3090 NVIDIA GeForce RTX 3090 Ti NVIDIA GeForce RTX 4090
envelope: only frontier
arms
click to show/hide · shift-click to compare · frontiers recompute over what is visible
all none clear comparison