[{"id":"4","card_a":"RTX 3090","vram_a_gb":"24","card_b":"RTX 3090","vram_b_gb":"24","link_class":"same-box","quant_family":"TQ1_0","quant_tier":"gyro-s","engine":"agention-llamacpp","engine_version":"dc255f2","draft_flag":"1","prefill_tps":"386.1","decode_tps":"82.4","cost_usd":"3000","currency_date":"2026-11-01","source_kind":"bench","verified_level":"bench-verified","harness_hash":"taichi-llamacli-v1","measured_at":"2026-11-01","notes":"split + MTP draft 4","active":"1"},{"id":"2","card_a":"RTX 3090","vram_a_gb":"24","card_b":"RTX 3090","vram_b_gb":"24","link_class":"same-box","quant_family":"TQ2_T","quant_tier":"gyro-coder","engine":"agention-llamacpp","engine_version":"dc255f2","draft_flag":"1","prefill_tps":"400.8","decode_tps":"68.8","cost_usd":"3000","currency_date":"2026-11-01","source_kind":"bench","verified_level":"bench-verified","harness_hash":"taichi-llamacli-v1","measured_at":"2026-11-01","notes":"split + MTP draft 4","active":"1"},{"id":"3","card_a":"RTX 3090","vram_a_gb":"24","card_b":"RTX 3090","vram_b_gb":"24","link_class":"same-box","quant_family":"TQ1_0","quant_tier":"gyro-s","engine":"agention-llamacpp","engine_version":"dc255f2","draft_flag":"0","prefill_tps":"405.6","decode_tps":"48.4","cost_usd":"3000","currency_date":"2026-11-01","source_kind":"bench","verified_level":"bench-verified","harness_hash":"taichi-llamacli-v1","measured_at":"2026-11-01","notes":"pool fits no single card; 13.9\/14.1 GiB per card","active":"1"},{"id":"1","card_a":"RTX 3090","vram_a_gb":"24","card_b":"RTX 3090","vram_b_gb":"24","link_class":"same-box","quant_family":"TQ2_T","quant_tier":"gyro-coder","engine":"agention-llamacpp","engine_version":"dc255f2","draft_flag":"0","prefill_tps":"418.0","decode_tps":"44.9","cost_usd":"3000","currency_date":"2026-11-01","source_kind":"bench","verified_level":"bench-verified","harness_hash":"taichi-llamacli-v1","measured_at":"2026-11-01","notes":"experts split 50\/50 across both cards","active":"1"},{"id":"5","card_a":"RTX 3090","vram_a_gb":"24","card_b":"RTX 3090","vram_b_gb":"24","link_class":"same-box","quant_family":"TQ2_0","quant_tier":"gyro-m","engine":"agention-llamacpp","engine_version":"dc255f2","draft_flag":"0","prefill_tps":"402.6","decode_tps":"44.5","cost_usd":"3000","currency_date":"2026-11-01","source_kind":"bench","verified_level":"bench-verified","harness_hash":"taichi-llamacli-v1","measured_at":"2026-11-01","notes":"split; 17.7\/18.1 GB per card","active":"1"},{"id":"9","card_a":"RTX 3090","vram_a_gb":"24","card_b":"RTX 3090","vram_b_gb":"24","link_class":"same-box","quant_family":"TQ2_T","quant_tier":"gyro-coder","engine":"agention-llamacpp","engine_version":"dc255f2","draft_flag":"0","prefill_tps":"515.0","decode_tps":"40.7","cost_usd":"3000","currency_date":"2026-11-01","source_kind":"bench","verified_level":"bench-verified","harness_hash":"taichi-llamacli-v1","measured_at":"2026-11-01","notes":"Vulkan host + CUDA experts \u2014 reversal is free","active":"1"},{"id":"7","card_a":"RTX 3090","vram_a_gb":"24","card_b":"V100-32G","vram_b_gb":"32","link_class":"rpc-bridge","quant_family":"TQ2_T","quant_tier":"gyro-coder","engine":"agention-llamacpp","engine_version":"dc255f2","draft_flag":"0","prefill_tps":"291.1","decode_tps":"33.6","cost_usd":"2500","currency_date":"2026-11-01","source_kind":"bench","verified_level":"bench-verified","harness_hash":"taichi-llamacli-v1","measured_at":"2026-11-01","notes":"experts on V100 via RPC over vnet bridge","active":"1"},{"id":"8","card_a":"RTX 3090","vram_a_gb":"24","card_b":"RTX 3090","vram_b_gb":"24","link_class":"same-box","quant_family":"TQ2_T","quant_tier":"gyro-coder","engine":"agention-llamacpp","engine_version":"dc255f2","draft_flag":"0","prefill_tps":"195.7","decode_tps":"22.0","cost_usd":"3000","currency_date":"2026-11-01","source_kind":"bench","verified_level":"bench-verified","harness_hash":"taichi-llamacli-v1","measured_at":"2026-11-01","notes":"experts on Vulkan\/RADV side \u2014 the 0.45x penalty row","active":"1"},{"id":"6","card_a":"RTX 3090","vram_a_gb":"24","card_b":null,"vram_b_gb":null,"link_class":"same-box","quant_family":"TQ2_T","quant_tier":"gyro-coder","engine":"strata","engine_version":"0.2.x","draft_flag":"0","prefill_tps":null,"decode_tps":"15.0","cost_usd":"1500","currency_date":"2026-11-01","source_kind":"bench","verified_level":"bench-verified","harness_hash":"taichi-strata-gcoder","measured_at":"2026-10-24","notes":"single card, experts VRAM-hot","active":"1"},{"id":"12","card_a":"R9700-32G","vram_a_gb":"32","card_b":null,"vram_b_gb":null,"link_class":"same-box","quant_family":"TQ1_0","quant_tier":"gyro-s","engine":"strata","engine_version":"rc2","draft_flag":"1","prefill_tps":null,"decode_tps":"89.7","cost_usd":"2400","currency_date":"2026-11-01","source_kind":"upstream_quote","verified_level":"self-reported","harness_hash":null,"measured_at":"2026-10-16","notes":"--spec 4 --mtp --kv int8; the number to beat","active":"1"},{"id":"10","card_a":"RTX 3090","vram_a_gb":"24","card_b":null,"vram_b_gb":null,"link_class":"same-box","quant_family":"TQ1_0","quant_tier":"gyro-s","engine":"strata","engine_version":"0.1.x","draft_flag":"1","prefill_tps":null,"decode_tps":"45.2","cost_usd":"1500","currency_date":"2026-11-01","source_kind":"upstream_quote","verified_level":"self-reported","harness_hash":null,"measured_at":"2026-10-03","notes":"their published best single-3090 JSON task; MTP draft on by default","active":"1"},{"id":"13","card_a":"V100-32G","vram_a_gb":"32","card_b":"V100-32G","vram_b_gb":"32","link_class":"nvlink","quant_family":"Q5_K_M","quant_tier":"qwen38-fn","engine":"llama.cpp","engine_version":"master","draft_flag":"0","prefill_tps":null,"decode_tps":"30.0","cost_usd":"2000","currency_date":"2026-11-01","source_kind":"upstream_quote","verified_level":"self-reported","harness_hash":null,"measured_at":"2026-10-28","notes":"reddit 2xV100 NVLink; 30 at ctx-0 -> 15 at 160k","active":"1"},{"id":"11","card_a":"Arc A750","vram_a_gb":"8","card_b":null,"vram_b_gb":null,"link_class":"same-box","quant_family":"TQ1_0","quant_tier":"gyro-s","engine":"strata","engine_version":"0.2.2","draft_flag":"1","prefill_tps":null,"decode_tps":"15.8","cost_usd":"150","currency_date":"2026-11-01","source_kind":"upstream_quote","verified_level":"self-reported","harness_hash":null,"measured_at":"2026-10-03","notes":"410\/24576 experts on-card; CPU-pool-bound by design","active":"1"}]