{"source": "https://homelaber.dev", "langue": "fr", "description": "mesures faites par l'auteur sur son propre matériel ; une ligne par mesure, avec ses conditions", "mesures": [
  {
    "mesure": "codes retrouvés dans un long prompt",
    "valeur": 3,
    "unite": "sur 3",
    "materiel": "RTX 3090 + RTX 3080 Ti, 80 Go DDR4",
    "modele": "Qwen3.8-Flash-Next 125B MoE IQ3_XXS",
    "moteur": "Strata",
    "conditions": "prompt de 373 723 tokens, codes à 5, 50 et 92% de profondeur, 400K de contexte par slot (YaRN x2)",
    "runs": 1,
    "experience": "https://homelaber.dev/labo/400k-contexte-80go/",
    "date": "2026-10-19"
  },
  {
    "mesure": "lecture du prompt (prefill)",
    "valeur": 1914,
    "unite": "tok/s",
    "materiel": "RTX 3090 + RTX 3080 Ti, 80 Go DDR4",
    "modele": "Qwen3.8-Flash-Next 125B MoE IQ3_XXS",
    "moteur": "Strata",
    "conditions": "prompt de 373 723 tokens lu en 195 s",
    "runs": 1,
    "experience": "https://homelaber.dev/labo/400k-contexte-80go/",
    "date": "2026-10-19"
  },
  {
    "mesure": "decode de la réponse",
    "valeur": 69,
    "unite": "tok/s",
    "materiel": "RTX 3090 + RTX 3080 Ti, 80 Go DDR4",
    "modele": "Qwen3.8-Flash-Next 125B MoE IQ3_XXS",
    "moteur": "Strata",
    "conditions": "après le prompt de 373 723 tokens",
    "runs": 1,
    "experience": "https://homelaber.dev/labo/400k-contexte-80go/",
    "date": "2026-10-19"
  },
  {
    "mesure": "cache K/V en ram par session",
    "valeur": 3.63,
    "unite": "GiB",
    "materiel": "RTX 3090 + RTX 3080 Ti, 80 Go DDR4",
    "modele": "Qwen3.8-Flash-Next 125B MoE IQ3_XXS",
    "moteur": "Strata",
    "conditions": "400K de contexte, K/V en int8 (1,16 à 131K, 2,32 à 262K)",
    "runs": 1,
    "experience": "https://homelaber.dev/labo/400k-contexte-80go/",
    "date": "2026-10-19"
  },
  {
    "mesure": "cache K/V en ram par session",
    "valeur": 1.16,
    "unite": "GiB par 131K tokens",
    "materiel": "RTX 3090 + RTX 3080 Ti, 80 Go DDR4",
    "modele": "Qwen3.8-Flash-Next 125B MoE IQ3_XXS",
    "moteur": "Strata",
    "conditions": "K/V en int8, mesuré",
    "runs": 1,
    "experience": "https://homelaber.dev/labo/512k-machine-figee/",
    "date": "2026-10-19"
  },
  {
    "mesure": "cache K/V estimé pour 4 slots + 1 session à 512K",
    "valeur": 23,
    "unite": "Go",
    "materiel": "RTX 3090 + RTX 3080 Ti, 80 Go DDR4",
    "modele": "Qwen3.8-Flash-Next 125B MoE IQ3_XXS",
    "moteur": "Strata",
    "conditions": "calcul à partir de la taille mesurée par session",
    "runs": 0,
    "experience": "https://homelaber.dev/labo/512k-machine-figee/",
    "date": "2026-10-19"
  },
  {
    "mesure": "decode, une seule requête",
    "valeur": 91.7,
    "unite": "tok/s",
    "materiel": "RTX 3090 + RTX 3080 Ti, 80 Go DDR4",
    "modele": "Qwen3.8-Flash-Next 125B MoE IQ3_XXS",
    "moteur": "Strata 0.1.45",
    "conditions": "2 slots, 262K de contexte, cartes à 220 W, réponses de 256 tokens, médiane de 3",
    "runs": 1,
    "experience": "https://homelaber.dev/labo/strata-020-ampere/",
    "date": "2026-10-19"
  },
  {
    "mesure": "decode, une seule requête",
    "valeur": 107.6,
    "unite": "tok/s",
    "materiel": "RTX 3090 + RTX 3080 Ti, 80 Go DDR4",
    "modele": "Qwen3.8-Flash-Next 125B MoE IQ3_XXS",
    "moteur": "Strata 0.2.0-beta",
    "conditions": "même config",
    "runs": 1,
    "experience": "https://homelaber.dev/labo/strata-020-ampere/",
    "date": "2026-10-19"
  },
  {
    "mesure": "decode par requête, 2 en même temps",
    "valeur": 61.1,
    "unite": "tok/s",
    "materiel": "RTX 3090 + RTX 3080 Ti, 80 Go DDR4",
    "modele": "Qwen3.8-Flash-Next 125B MoE IQ3_XXS",
    "moteur": "Strata 0.2.0-beta",
    "conditions": "même config (0.1.45 : 54,7)",
    "runs": 1,
    "experience": "https://homelaber.dev/labo/strata-020-ampere/",
    "date": "2026-10-19"
  },
  {
    "mesure": "temps total d'une charge de 4 agents",
    "valeur": 215,
    "unite": "s",
    "materiel": "RTX 3090 + RTX 3080 Ti, 80 Go DDR4",
    "modele": "Qwen3.8-Flash-Next 125B MoE IQ3_XXS",
    "moteur": "Strata 0.2.0-beta",
    "conditions": "vraie charge d'agents de code sur un repo, contexte jusqu'à 122K (0.1.45 : 212 s)",
    "runs": 1,
    "experience": "https://homelaber.dev/labo/strata-020-ampere/",
    "date": "2026-10-19"
  },
  {
    "mesure": "decode par agent, deux requêtes en même temps",
    "valeur": 43.8,
    "unite": "tok/s",
    "materiel": "RTX 3090 + RTX 3080 Ti, 80 Go DDR4",
    "modele": "Qwen3.8-Flash-Next 125B MoE IQ3_XXS",
    "moteur": "Strata 0.1.45",
    "conditions": "batching par défaut, 4 slots, cartes à 220 W, réponses de 256 tokens, médiane de 3 manches",
    "runs": 1,
    "experience": "https://homelaber.dev/labo/deux-agents-44-66-tok-s/",
    "date": "2026-10-20"
  },
  {
    "mesure": "decode par agent, deux requêtes en même temps",
    "valeur": 52.3,
    "unite": "tok/s",
    "materiel": "RTX 3090 + RTX 3080 Ti, 80 Go DDR4",
    "modele": "Qwen3.8-Flash-Next 125B MoE IQ3_XXS",
    "moteur": "Strata 0.1.45 + PR #2985",
    "conditions": "2 slots, drafts MTP dans les slots batchés",
    "runs": 1,
    "experience": "https://homelaber.dev/labo/deux-agents-44-66-tok-s/",
    "date": "2026-10-20"
  },
  {
    "mesure": "decode par agent, deux requêtes en même temps",
    "valeur": 61.1,
    "unite": "tok/s",
    "materiel": "RTX 3090 + RTX 3080 Ti, 80 Go DDR4",
    "modele": "Qwen3.8-Flash-Next 125B MoE IQ3_XXS",
    "moteur": "Strata 0.2.0-beta + PR #2985",
    "conditions": "2 slots, cartes à 220 W",
    "runs": 1,
    "experience": "https://homelaber.dev/labo/deux-agents-44-66-tok-s/",
    "date": "2026-10-20"
  },
  {
    "mesure": "decode par agent, deux requêtes en même temps",
    "valeur": 66.4,
    "unite": "tok/s",
    "materiel": "RTX 3090 + RTX 3080 Ti, 80 Go DDR4",
    "modele": "Qwen3.8-Flash-Next 125B MoE IQ3_XXS",
    "moteur": "Strata 0.2.0-beta + PR #2985",
    "conditions": "2 slots, cartes à 260 W, calibration auto",
    "runs": 1,
    "experience": "https://homelaber.dev/labo/deux-agents-44-66-tok-s/",
    "date": "2026-10-20"
  },
  {
    "mesure": "decode, une seule requête",
    "valeur": 109.6,
    "unite": "tok/s",
    "materiel": "RTX 3090 + RTX 3080 Ti, 80 Go DDR4",
    "modele": "Qwen3.8-Flash-Next 125B MoE IQ3_XXS",
    "moteur": "Strata 0.2.0-beta + PR #2985",
    "conditions": "2 slots, cartes à 260 W",
    "runs": 1,
    "experience": "https://homelaber.dev/labo/deux-agents-44-66-tok-s/",
    "date": "2026-10-20"
  },
  {
    "mesure": "decode moyen (8K / 32K / 120K)",
    "valeur": 26.1,
    "unite": "tok/s",
    "materiel": "RTX 3090 + RTX 3080 Ti, 78 Go DDR4",
    "modele": "GLM-5.3-Flash 321B, Maya-S24 IQ2_XXS_S",
    "moteur": "Project Maya v1.1.11",
    "conditions": "MTP, KV int8, 512 tokens générés",
    "runs": 2,
    "experience": "https://homelaber.dev/labo/glm-53-maya-3090-3080ti/",
    "date": "2026-10-20"
  },
  {
    "mesure": "decode moyen (8K / 32K / 120K)",
    "valeur": 31.7,
    "unite": "tok/s",
    "materiel": "RTX 3090 + RTX 3080 Ti, 78 Go DDR4",
    "modele": "GLM-5.3-Flash 321B, Maya-S24 IQ2_XXS_S",
    "moteur": "Project Maya v1.1.18",
    "conditions": "STRATA_GLM_ROUTE_AFFINITY=0.02 et STRATA_GLM_DEFER_CPU=1",
    "runs": 2,
    "experience": "https://homelaber.dev/labo/glm-53-maya-3090-3080ti/",
    "date": "2026-10-20"
  },
  {
    "mesure": "étapes d'agent en 5 minutes",
    "valeur": 24,
    "unite": "étapes",
    "materiel": "RTX 3090 + RTX 3080 Ti, 78 Go DDR4",
    "modele": "GLM-5.3-Flash 321B",
    "moteur": "Project Maya v1.1.18",
    "conditions": "review en lecture seule d'un repo TypeScript (Qwen3.8-Flash-Next sur Strata : 36)",
    "runs": 1,
    "experience": "https://homelaber.dev/labo/glm-53-maya-3090-3080ti/",
    "date": "2026-10-20"
  },
  {
    "mesure": "decode à vide",
    "valeur": 136.2,
    "unite": "tok/s",
    "materiel": "RTX 3080 Ti 12 Go (260 W)",
    "modele": "Spark-X2.6-4B Q8_0",
    "moteur": "llama.cpp (commit 839a572)",
    "conditions": "llama-bench tg128, KV q8_0, flash attention",
    "runs": 2,
    "experience": "https://homelaber.dev/labo/spark-4b-256k-3080ti/",
    "date": "2026-10-20"
  },
  {
    "mesure": "decode à 254K de contexte",
    "valeur": 34.9,
    "unite": "tok/s",
    "materiel": "RTX 3080 Ti 12 Go (260 W)",
    "modele": "Spark-X2.6-4B Q8_0",
    "moteur": "llama.cpp (commit 839a572)",
    "conditions": "llama-bench, profondeur 254K, KV q8_0",
    "runs": 2,
    "experience": "https://homelaber.dev/labo/spark-4b-256k-3080ti/",
    "date": "2026-10-20"
  },
  {
    "mesure": "decode à 128K de contexte, KV f16",
    "valeur": 78.2,
    "unite": "tok/s",
    "materiel": "RTX 3080 Ti 12 Go (260 W)",
    "modele": "Spark-X2.6-4B Q8_0",
    "moteur": "llama.cpp (commit 839a572)",
    "conditions": "llama-bench, KV f16 (q8_0 : 57,4)",
    "runs": 2,
    "experience": "https://homelaber.dev/labo/spark-4b-256k-3080ti/",
    "date": "2026-10-20"
  },
  {
    "mesure": "codes retrouvés, multi-needle",
    "valeur": 40,
    "unite": "sur 40",
    "materiel": "RTX 3080 Ti 12 Go",
    "modele": "Spark-X2.6-4B Q8_0",
    "moteur": "llama-server",
    "conditions": "4 codes par prompt, prompts de 15,8K à ~245K tokens, 2 essais par longueur",
    "runs": 10,
    "experience": "https://homelaber.dev/labo/spark-4b-256k-3080ti/",
    "date": "2026-10-20"
  },
  {
    "mesure": "decode par agent, deux agents en même temps",
    "valeur": 67,
    "unite": "tok/s",
    "materiel": "RTX 3090 + RTX 3080 Ti, 80 Go DDR4",
    "modele": "Qwen3.8-Flash-Next 125B MoE IQ3_XXS",
    "moteur": "Strata 0.2.0-beta",
    "conditions": "les deux cartes à 260 W, d'origine",
    "runs": 1,
    "experience": "https://homelaber.dev/labo/undervolt-3090-3080ti/",
    "date": "2026-10-20"
  },
  {
    "mesure": "decode par agent, deux agents en même temps",
    "valeur": 56.8,
    "unite": "tok/s",
    "materiel": "RTX 3090 + RTX 3080 Ti, 80 Go DDR4",
    "modele": "Qwen3.8-Flash-Next 125B MoE IQ3_XXS",
    "moteur": "Strata 0.2.0-beta",
    "conditions": "les deux cartes limitées à 220 W",
    "runs": 1,
    "experience": "https://homelaber.dev/labo/undervolt-3090-3080ti/",
    "date": "2026-10-20"
  },
  {
    "mesure": "decode par agent, deux agents en même temps",
    "valeur": 63.5,
    "unite": "tok/s",
    "materiel": "RTX 3090 + RTX 3080 Ti, 80 Go DDR4",
    "modele": "Qwen3.8-Flash-Next 125B MoE IQ3_XXS",
    "moteur": "Strata 0.2.0-beta",
    "conditions": "3090 bloquée à 1500 MHz + offset +150 MHz, 3080 Ti à 260 W",
    "runs": 1,
    "experience": "https://homelaber.dev/labo/undervolt-3090-3080ti/",
    "date": "2026-10-20"
  },
  {
    "mesure": "decode par agent, deux agents en même temps",
    "valeur": 62,
    "unite": "tok/s",
    "materiel": "RTX 3090 + RTX 3080 Ti, 80 Go DDR4",
    "modele": "Qwen3.8-Flash-Next 125B MoE IQ3_XXS",
    "moteur": "Strata 0.2.0-beta",
    "conditions": "3090 bloquée à 1500 MHz + 150, 3080 Ti bloquée à 1800 MHz + 150",
    "runs": 1,
    "experience": "https://homelaber.dev/labo/undervolt-3090-3080ti/",
    "date": "2026-10-20"
  },
  {
    "mesure": "conso des deux cartes",
    "valeur": 452,
    "unite": "W",
    "materiel": "RTX 3090 + RTX 3080 Ti",
    "conditions": "d'origine, 260 W chacune, deux agents",
    "runs": 1,
    "experience": "https://homelaber.dev/labo/undervolt-3090-3080ti/",
    "date": "2026-10-20"
  },
  {
    "mesure": "conso des deux cartes",
    "valeur": 358,
    "unite": "W",
    "materiel": "RTX 3090 + RTX 3080 Ti",
    "conditions": "les deux cartes undervoltées, deux agents",
    "runs": 1,
    "experience": "https://homelaber.dev/labo/undervolt-3090-3080ti/",
    "date": "2026-10-20"
  }
]}