{"model_id":"Qwen/Qwen3-8B","runtime":"vllm","runtime_version":"0.27.1","gpu":"RTX PRO 6000 Blackwell","gpu_arch":"sm_120","vram_gb":96,"quantization":"bf16","tp":1,"backend":null,"context_len":null,"concurrency":32,"startup_status":"ok","correctness_status":"greedy-matched","agg_tokens_per_second":1725,"ttft_ms_p50":39,"e2e_s_p99":3.4,"peak_vram_gb":null,"artifact_url":"https://conatus.jahn.ai/ai-engineering/sample-report/","tested_at":"2026-08-25"}
{"model_id":"Qwen/Qwen3-8B","runtime":"sglang","runtime_version":"0.5.9","gpu":"RTX PRO 6000 Blackwell","gpu_arch":"sm_120","vram_gb":96,"quantization":"bf16","tp":1,"backend":null,"context_len":null,"concurrency":32,"startup_status":"ok","correctness_status":"greedy-matched","agg_tokens_per_second":1327,"ttft_ms_p50":42,"e2e_s_p99":5.0,"peak_vram_gb":null,"artifact_url":"https://conatus.jahn.ai/ai-engineering/sample-report/","tested_at":"2026-08-25"}
{"model_id":"Qwen/Qwen3-8B","runtime":"llama.cpp","runtime_version":"cuda","gpu":"RTX PRO 6000 Blackwell","gpu_arch":"sm_120","vram_gb":96,"quantization":"bf16","tp":1,"backend":null,"context_len":null,"concurrency":32,"startup_status":"ok","correctness_status":"greedy-matched","agg_tokens_per_second":428,"ttft_ms_p50":316,"e2e_s_p99":16.3,"peak_vram_gb":null,"artifact_url":"https://conatus.jahn.ai/ai-engineering/sample-report/","tested_at":"2026-08-25"}
{"model_id":"Qwen/Qwen3-8B","runtime":"vllm","runtime_version":"0.27.1","gpu":"RTX PRO 6000 Blackwell","gpu_arch":"sm_120","vram_gb":96,"quantization":"fp8","tp":1,"backend":null,"context_len":null,"concurrency":32,"startup_status":"ok","correctness_status":"no-regression-20prompt","agg_tokens_per_second":2597,"ttft_ms_p50":null,"e2e_s_p99":null,"peak_vram_gb":null,"artifact_url":"https://conatus.jahn.ai/ai-engineering/sample-report/","tested_at":"2026-08-25"}
{"model_id":"nvidia/Qwen3.6-35B-A3B-NVFP4","runtime":"vllm","runtime_version":"0.25.1","gpu":"RTX PRO 6000 Blackwell","gpu_arch":"sm_120","vram_gb":96,"quantization":"nvfp4","kv_cache_dtype":"fp8","tp":1,"backend":"flashinfer_b12x","context_len":180224,"concurrency":null,"startup_status":"ok","correctness_status":"not-audited","available_kv_cache_gib":52.89,"max_concurrency_x":29.73,"peak_vram_gb":91.8,"artifact_url":"https://conatus.jahn.ai/ai-engineering/sm120-b12x-workspace/","tested_at":"2026-08-26","note":"OOMs on 16-32GB Blackwell during profile_run; starts on 96GB"}
{"model_id":"nvidia/Qwen3.6-35B-A3B-NVFP4","runtime":"vllm","runtime_version":"0.25.1","gpu":"RTX PRO 6000 Blackwell","gpu_arch":"sm_120","vram_gb":96,"quantization":"nvfp4","kv_cache_dtype":"fp8","tp":1,"backend":"marlin","context_len":180224,"concurrency":null,"startup_status":"ok","correctness_status":"not-audited","available_kv_cache_gib":64.36,"max_concurrency_x":36.17,"peak_vram_gb":88.97,"artifact_url":"https://conatus.jahn.ai/ai-engineering/sm120-b12x-workspace/","tested_at":"2026-08-26","note":"marlin baseline; 11.47 GiB more KV headroom than flashinfer_b12x"}
{"model_id":"unsloth/Qwen3.8-27B-NVFP4","runtime":"sglang","runtime_version":"c7e2c08d1 (pre-#35228; all releases <=0.5.18)","gpu":"RTX PRO 6000 Blackwell","gpu_arch":"sm_120","vram_gb":96,"quantization":"nvfp4+fp8-lm_head (compressed-tensors mixed)","tp":1,"backend":"flashinfer","context_len":null,"concurrency":1,"startup_status":"ok","correctness_status":"FAIL: degenerate repetition, empty content (FP8 lm_head weight_scale never applied, sglang#34895)","agg_tokens_per_second":null,"ttft_ms_p50":null,"e2e_s_p99":null,"peak_vram_gb":null,"artifact_url":"https://github.com/sgl-project/sglang/issues/34895#issuecomment-5420022002","tested_at":"2026-08-26"}
{"model_id":"unsloth/Qwen3.8-27B-NVFP4","runtime":"sglang","runtime_version":"main 07a9de25b (>=5375babb, PR #35228)","gpu":"RTX PRO 6000 Blackwell","gpu_arch":"sm_120","vram_gb":96,"quantization":"nvfp4+fp8-lm_head (compressed-tensors mixed)","tp":1,"backend":"flashinfer","context_len":null,"concurrency":1,"startup_status":"ok","correctness_status":"greedy spot-check correct (lm_head weight_scale loads, 0 warnings)","agg_tokens_per_second":null,"ttft_ms_p50":null,"e2e_s_p99":null,"peak_vram_gb":null,"artifact_url":"https://github.com/sgl-project/sglang/issues/34895#issuecomment-5420022002","tested_at":"2026-08-26"}