{ "date": "2026-10-06", "metric": "Inverse median HTTP TPOT: (last nonempty delta time - first)/(completion tokens - first delta tokens). Reasoning included.", "conditions": { "hardware": "Mac mini M4 Pro, CPU12/GPU16, 64GB, macOS27.0.1", "ollama": "0.34.0, qwen3.8:27b-mlx, NVFP4", "splash": "1.2.1, incoai/Qwen3.8-27B-Splash, affine Q4 + DFlash2", "reasoning_effort": "medium", "http_max_tokens": 1024, "http_window": 131072, "agent_window": 32768, "http_requests_per_bucket": 3, "128k_distinct_inputs": 1, "agent_repeats": 1, "agent_response_budget": "Pi default", "agent_default_model_max_tokens": 16384, "agent_request_output_budgets": { "ollama": [ 16384, 16384, 16384, 16384, 16384, 16384, 16384, 16384, 16384, 16384, 16384, 16268, 15955, 15452, 14479, 7051, 6313, 5619, 5050, 4509 ], "splash": [ 16384, 16384, 16384, 16384, 16384, 16384, 16384, 16384, 16384, 16384, 16384, 16384, 16384, 16384, 16384, 16384, 16384, 16152, 15659, 15072 ] } }, "vendor": { "source": "https://inco.ai/blog/splash/", "hardware": "M5 Pro / 48GB", "model": "Qwen3.8-27B", "metric": "Decode throughput from inverse P50 TPOT, reasoning included", "buckets_k": [ 8, 16, 32 ], "series": { "Splash": [ 55, 55, 54 ], "oMLX": [ 33, 29, 28 ], "Ollama": [ 21, 21, 19 ], "uzu": [ 18, 17, 15 ] } }, "curve": { "metric": "Inverse median HTTP TPOT: (last nonempty delta time - first)/(completion tokens - first delta tokens). Reasoning included.", "reasoning_effort": "medium", "max_tokens": 1024, "window": 131072, "rows": [ { "bucket_k": 8, "engines": { "ollama": { "decode_tps": 22.02097446530405, "decode_tpot_s": 0.04541125105864804, "decode_tps_min": 20.915252684161985, "decode_tps_max": 23.129842287634034, "requests": 3, "distinct_prompts": 3, "prompt_tokens": [ 8161, 8197, 8196 ], "output_tokens": [ 1024, 1024, 1024 ], "first_chunk_tokens": [ 1, 1, 1 ], "ttft_s": [ 88.86414562500431, 88.50570270900789, 93.67371237499174 ], "total_s": [ 135.3201025830058, 137.41774408399942, 137.90383033300168 ], "cache_tokens": [ 0, 0, 2 ], "cases": [ "8k_0_r0", "8k_1_r0", "8k_2_r0" ], "server_decode_tps": [ 21.845009483805644, 20.887568080211448, 23.10262572424525 ], "finish_reasons": [ "length", "length", "length" ], "reasoning_observed": [ true, true, true ] }, "splash": { "decode_tps": 36.78882209383913, "decode_tpot_s": 0.027182169558167664, "decode_tps_min": 32.79148354063038, "decode_tps_max": 40.289092815661185, "requests": 3, "distinct_prompts": 3, "prompt_tokens": [ 8201, 8197, 8197 ], "output_tokens": [ 1024, 1024, 1024 ], "first_chunk_tokens": [ 1, 1, 1 ], "ttft_s": [ 83.03516129199124, 85.29060879099416, 87.0460452499974 ], "total_s": [ 114.23268495799857, 110.68362933299795, 114.85365575000469 ], "cache_tokens": [ 0, 0, 0 ], "cases": [ "8k_0_r0", "8k_1_r0", "8k_2_r0" ], "server_decode_tps": [ 32.56755764794278, 40.017122286811535, 36.53720771582395 ], "finish_reasons": [ "length", "length", "length" ], "reasoning_observed": [ true, true, true ] } }, "decode_speed_increase_pct": 67.06264362552665 }, { "bucket_k": 16, "engines": { "ollama": { "decode_tps": 21.019657633129672, "decode_tpot_s": 0.047574514174002144, "decode_tps_min": 18.97292703853771, "decode_tps_max": 22.34745649119835, "requests": 3, "distinct_prompts": 3, "prompt_tokens": [ 16366, 16382, 16362 ], "output_tokens": [ 1024, 1024, 1024 ], "first_chunk_tokens": [ 1, 1, 1 ], "ttft_s": [ 196.79365083300218, 193.09759716599365, 193.78922400000738 ], "total_s": [ 242.57096650000312, 247.01699508300226, 242.45827808401373 ], "cache_tokens": [ 2, 5, 9 ], "cases": [ "16k_0_r0", "16k_1_r0", "16k_2_r0" ], "server_decode_tps": [ 22.232629634590275, 18.84089872616422, 20.739638462502345 ], "finish_reasons": [ "length", "length", "length" ], "reasoning_observed": [ true, true, true ] }, "splash": { "decode_tps": 35.48434318890306, "decode_tpot_s": 0.02818144314173829, "decode_tps_min": 33.225851250142114, "decode_tps_max": 42.013719347415254, "requests": 3, "distinct_prompts": 3, "prompt_tokens": [ 16393, 16392, 16391 ], "output_tokens": [ 1024, 1024, 1024 ], "first_chunk_tokens": [ 1, 1, 1 ], "ttft_s": [ 175.2467440420005, 183.095292375001, 175.5539672080049 ], "total_s": [ 199.5961437090009, 213.8848916660063, 204.38386304200685 ], "cache_tokens": [ 0, 0, 0 ], "cases": [ "16k_0_r0", "16k_1_r0", "16k_2_r0" ], "server_decode_tps": [ 41.72685608222523, 33.03152308352938, 35.27662440790683 ], "finish_reasons": [ "length", "length", "length" ], "reasoning_observed": [ true, true, true ] } }, "decode_speed_increase_pct": 68.81503879956253 }, { "bucket_k": 32, "engines": { "ollama": { "decode_tps": 18.238408856856744, "decode_tpot_s": 0.05482934437145537, "decode_tps_min": 16.61171301068048, "decode_tps_max": 18.624031589594345, "requests": 3, "distinct_prompts": 3, "prompt_tokens": [ 32740, 32748, 32603 ], "output_tokens": [ 1024, 1024, 1024 ], "first_chunk_tokens": [ 1, 1, 1 ], "ttft_s": [ 443.5855730829935, 580.6548085420072, 550.2957565830002 ], "total_s": [ 505.1694154579891, 636.7457016250119, 605.2267312910117 ], "cache_tokens": [ 5, 5, 9 ], "cases": [ "32k_0_r0", "32k_1_r0", "32k_2_r0" ], "server_decode_tps": [ 16.504365634296637, 18.112947053408128, 18.347466328563833 ], "finish_reasons": [ "length", "length", "length" ], "reasoning_observed": [ true, true, true ] }, "splash": { "decode_tps": 31.70880534287943, "decode_tpot_s": 0.03153698126393025, "decode_tps_min": 25.91672807274788, "decode_tps_max": 34.10928985496996, "requests": 3, "distinct_prompts": 3, "prompt_tokens": [ 32775, 32777, 32776 ], "output_tokens": [ 557, 1024, 1024 ], "first_chunk_tokens": [ 1, 1, 1 ], "ttft_s": [ 378.82640858300147, 365.9870400420041, 362.9404434169992 ], "total_s": [ 400.28000912499556, 395.9791685829987, 395.2030317079916 ], "cache_tokens": [ 0, 0, 0 ], "cases": [ "32k_0_r0", "32k_1_r0", "32k_2_r0" ], "server_decode_tps": [ 25.59078709290197, 33.87648068312057, 31.585136821915313 ], "finish_reasons": [ "stop", "length", "length" ], "reasoning_observed": [ true, true, true ] } }, "decode_speed_increase_pct": 73.85730077521802 }, { "bucket_k": 128, "engines": { "ollama": { "decode_tps": 8.920180767462034, "decode_tpot_s": 0.11210535145741442, "decode_tps_min": 7.8561634543151, "decode_tps_max": 10.409524392525274, "requests": 3, "distinct_prompts": 1, "prompt_tokens": [ 129571, 129571, 129571 ], "output_tokens": [ 401, 670, 1024 ], "first_chunk_tokens": [ 1, 1, 1 ], "ttft_s": [ 2272.4614555419976, 0.6137366249895422, 0.28854091600805987 ], "total_s": [ 2323.3821163749963, 75.86563000000024, 98.56423054100014 ], "cache_tokens": [ 5, 129566, 129570 ], "cases": [ "128k_0_r0", "128k_0_r1", "128k_0_r2" ], "server_decode_tps": [ 7.560878874438793, 8.872983548449447, 10.395245375160789 ], "finish_reasons": [ "stop", "stop", "length" ], "reasoning_observed": [ true, true, true ] }, "splash": { "decode_tps": 24.782915684087527, "decode_tpot_s": 0.04035037736266336, "decode_tps_min": 24.298447570583004, "decode_tps_max": 24.818964697390626, "requests": 3, "distinct_prompts": 1, "prompt_tokens": [ 129606, 129606, 129606 ], "output_tokens": [ 1024, 1024, 1024 ], "first_chunk_tokens": [ 1, 1, 1 ], "ttft_s": [ 2005.6259153749997, 0.7300170829985291, 0.6846968329919036 ], "total_s": [ 2047.728956167004, 42.008792000007816, 41.90349679099745 ], "cache_tokens": [ 0, 129600, 129600 ], "cases": [ "128k_0_r0", "128k_0_r1", "128k_0_r2" ], "server_decode_tps": [ 24.27385833477009, 24.759051668240435, 24.794849910732474 ], "finish_reasons": [ "length", "length", "length" ], "reasoning_observed": [ true, true, true ] } }, "decode_speed_increase_pct": 177.82974729041007 } ], "source_cases_sha256": "93d89a13caf51227fc513ac4485f041ab811a2bf1f97e5bdc956857da7948ff0", "source": "NVIDIA SPEED-Bench code_completion, existing trailing padding resized for Qwen tokenization; 128K adds padding.", "128k_scope": "One input, first cold request plus two exact replays; three decode observations, one cold TTFT.", "uncertainty": "Observed range of three requests, not a confidence interval; sequential engines, no ABBA." }, "agent": { "pi": "1.0.4", "window": 32768, "repeats": 1, "reasoning_effort": "medium", "max_tokens": null, "response_budget": "Pi default", "default_model_max_tokens": 16384, "initial_request_identical_except_model": true, "initial_request_sha256_except_model": "a8f322fb7677165255553ffaf07fe60c446bd20eec6786862f6aca003579669c", "max_prompt_tokens": { "ollama": 24164, "splash": 13588 }, "request_output_budgets": { "ollama": [ 16384, 16384, 16384, 16384, 16384, 16384, 16384, 16384, 16384, 16384, 16384, 16268, 15955, 15452, 14479, 7051, 6313, 5619, 5050, 4509 ], "splash": [ 16384, 16384, 16384, 16384, 16384, 16384, 16384, 16384, 16384, 16384, 16384, 16384, 16384, 16384, 16384, 16384, 16384, 16152, 15659, 15072 ] }, "tasks": [ { "task": "quote", "ollama_seconds": 111.42451174999587, "splash_seconds": 99.36780733299383, "ollama_success": true, "splash_success": true, "tests": 8, "waiting_time_reduction_pct": 10.820513572501689, "ollama_calls": 7, "splash_calls": 7, "ollama_tools": 7, "splash_tools": 8 }, { "task": "events", "ollama_seconds": 310.2816478340101, "splash_seconds": 110.39632716700726, "ollama_success": true, "splash_success": true, "tests": 18, "waiting_time_reduction_pct": 64.42060691063963, "ollama_calls": 6, "splash_calls": 7, "ollama_tools": 6, "splash_tools": 7 }, { "task": "csv", "ollama_seconds": 686.0503511660063, "splash_seconds": 129.8996844170033, "ollama_success": true, "splash_success": true, "tests": 27, "waiting_time_reduction_pct": 81.06557569772733, "ollama_calls": 7, "splash_calls": 6, "ollama_tools": 7, "splash_tools": 6 } ], "engines": { "ollama": { "success": true, "seconds": 1107.7565107500122, "tasks_completed": 3, "model_calls": 20, "tool_completed": 20, "blocked_tools": 0, "tool_errors": 2, "thinking_blocks": 20, "output_tokens": 13664, "truncated_final_responses": 0, "failures": [] }, "splash": { "success": true, "seconds": 339.6638189170044, "tasks_completed": 3, "model_calls": 20, "tool_completed": 21, "blocked_tools": 0, "tool_errors": 0, "thinking_blocks": 20, "output_tokens": 6031, "truncated_final_responses": 0, "failures": [] } }, "total_waiting_time_reduction_pct": 69.33768245812131 }, "source_curve_sha256": "50fd349d4d4c6e82b71cde409575dd01433b9d20cc33193b42d7ed3e1b08b167", "source_agent_sha256": "33ab34a66e6d7af894c20257aef93f3499057cdeb9999e2289a51e2441d27b90" }