Repository Public
1{2 "meta": {3 "date": "2026-11-19",4 "host": "Mac mini (m4mm)",5 "chip": "Apple M4 Pro",6 "memory_gb": 48,7 "macos": "26.5.1",8 "mlx": "0.32.4",9 "mlx_lm": "0.32.2",10 "model": "mlx-community/Qwen3-30B-A3B-Instruct-2507-4bit",11 "revision": "e9675aa3ca5f900ccef55267914466d55ab325fa"12 },13 "single": {14 "ttft_ms": 110.7,15 "tokens_per_second": 80.616 },17 "scenarios": [18 {19 "name": "short_prompt",20 "description": "Question courte, réponse de 128 tokens.",21 "requests": 5,22 "ttft_ms_p50": 110.7,23 "ttft_ms_p95": 146.3,24 "tokens_per_second_p50": 80.6,25 "latency_ms_p50": 1687.1,26 "latency_ms_p95": 1721.0,27 "error_rate": 0.028 },29 {30 "name": "long_generation",31 "description": "Génération longue de 1 024 tokens.",32 "requests": 3,33 "ttft_ms_p50": 140.9,34 "ttft_ms_p95": 187.9,35 "tokens_per_second_p50": 76.3,36 "latency_ms_p50": 13539.1,37 "latency_ms_p95": 13590.7,38 "error_rate": 0.039 },40 {41 "name": "long_prompt",42 "description": "Prompt d'environ 4 000 tokens (jamais vu), réponse de 128 tokens.",43 "requests": 3,44 "ttft_ms_p50": 4646.0,45 "ttft_ms_p95": 4650.4,46 "tokens_per_second_p50": 63.4,47 "latency_ms_p50": 6645.0,48 "latency_ms_p95": 6653.1,49 "error_rate": 0.050 },51 {52 "name": "large_context",53 "description": "Contexte d'environ 16 000 tokens (jamais vu), réponse de 128 tokens.",54 "requests": 2,55 "ttft_ms_p50": 27428.6,56 "ttft_ms_p95": 27637.8,57 "tokens_per_second_p50": 42.9,58 "latency_ms_p50": 30394.3,59 "latency_ms_p95": 30507.2,60 "error_rate": 0.061 },62 {63 "name": "coding",64 "description": "Programmation : fonction Python avec tests.",65 "requests": 3,66 "ttft_ms_p50": 83.2,67 "ttft_ms_p95": 338.3,68 "tokens_per_second_p50": 78.1,69 "latency_ms_p50": 6640.3,70 "latency_ms_p95": 6885.3,71 "error_rate": 0.072 },73 {74 "name": "reasoning",75 "description": "Raisonnement en plusieurs étapes.",76 "requests": 3,77 "ttft_ms_p50": 242.7,78 "ttft_ms_p95": 256.3,79 "tokens_per_second_p50": 78.2,80 "latency_ms_p50": 5339.9,81 "latency_ms_p95": 5344.1,82 "error_rate": 0.083 },84 {85 "name": "tool_calling",86 "description": "Appel d'outil (get_weather) attendu.",87 "requests": 3,88 "ttft_ms_p50": 453.3,89 "ttft_ms_p95": 586.2,90 "tokens_per_second_p50": null,91 "latency_ms_p50": 465.3,92 "latency_ms_p95": 598.1,93 "error_rate": 0.094 },95 {96 "name": "multi_turn",97 "description": "Conversation de 5 tours (préfixe réutilisé d'un tour à l'autre).",98 "requests": 5,99 "ttft_ms_p50": 118.9,100 "ttft_ms_p95": 170.0,101 "tokens_per_second_p50": 77.8,102 "latency_ms_p50": 1937.2,103 "latency_ms_p95": 2136.6,104 "error_rate": 0.0105 },106 {107 "name": "repeated_prefix",108 "description": "Préfixe système commun d'environ 3 000 tokens, 5 questions différentes après un premier appel à froid.",109 "requests": 5,110 "ttft_ms_p50": 147.2,111 "ttft_ms_p95": 149.5,112 "tokens_per_second_p50": 69.9,113 "latency_ms_p50": 391.1,114 "latency_ms_p95": 391.7,115 "error_rate": 0.0116 }117 ],118 "concurrency": [119 {120 "clients": 1,121 "aggregate_tps": 75.1,122 "per_request_tps": 78.7,123 "ttft_p50_ms": 145.9,124 "ttft_p95_ms": 204.9,125 "latency_p95_ms": 3427.6,126 "error_rate": 0.0127 },128 {129 "clients": 2,130 "aggregate_tps": 102.8,131 "per_request_tps": 53.4,132 "ttft_p50_ms": 194.4,133 "ttft_p95_ms": 196.3,134 "latency_p95_ms": 4979.2,135 "error_rate": 0.0136 },137 {138 "clients": 4,139 "aggregate_tps": 125.6,140 "per_request_tps": 32.6,141 "ttft_p50_ms": 321.6,142 "ttft_p95_ms": 330.2,143 "latency_p95_ms": 8155.2,144 "error_rate": 0.0145 },146 {147 "clients": 8,148 "aggregate_tps": 155.8,149 "per_request_tps": 20.4,150 "ttft_p50_ms": 547.5,151 "ttft_p95_ms": 695.1,152 "latency_p95_ms": 13278.9,153 "error_rate": 0.0154 },155 {156 "clients": 16,157 "aggregate_tps": 156.8,158 "per_request_tps": 20.5,159 "ttft_p50_ms": 13280.4,160 "ttft_p95_ms": 13924.6,161 "latency_p95_ms": 26152.4,162 "error_rate": 0.0163 }164 ],165 "peak_memory_bytes": 33855902318,166 "prefix_cache": {167 "cold_ttft_ms": 3564.2,168 "warm_ttft_ms": 147.2169 },170 "recommended": {171 "max_active": 8172 },173 "notes": [174 "Measured through the production API gateway on the serving machine, temperature 0.",175 "The engine decodes up to 8 requests in one continuous batch; additional requests wait in the priority queue, which is why time-to-first-token rises beyond 8 clients.",176 "Time to first token is measured client-side from request send to first content delta; tokens/s excludes the first token.",177 "Large prompts are dominated by prefill; a repeated prefix is served from the KV prefix cache."178 ]179}