{"id":33194,"date":"2026-06-27T17:05:00","date_gmt":"2026-06-27T15:05:00","guid":{"rendered":"https:\/\/www.chinasmartbuy.com\/blog\/?p=33194"},"modified":"2026-06-27T15:32:16","modified_gmt":"2026-06-27T13:32:16","slug":"dspark-deepseek-inferenza-ai","status":"publish","type":"post","link":"https:\/\/www.chinasmartbuy.com\/blog\/dspark-deepseek-inferenza-ai\/","title":{"rendered":"DSpark di DeepSeek accelera l&#8217;inferenza AI fino all&#8217;85%"},"content":{"rendered":"\n\n<style>\n  :root {\n    --red: #CC1B1B;\n    --red-light: rgba(204,27,27,0.08);\n    --red-mid: #e84040;\n    --accent-soft-bg: rgba(204,27,27,0.07);\n    --accent-soft-border: rgba(204,27,27,0.18);\n    --text: #1a1a1a;\n    --text-muted: #666;\n    --card: #ffffff;\n    --border: #ebebeb;\n    --page-bg: #ffffff;\n    --surface-soft: #fafafa;\n    --surface-contrast: #f0f0f0;\n    --line-soft: #f5f5f5;\n    --soft-label: #bbb;\n    --muted-label: #999;\n    --copy: #333;\n    --font-head: 'Bebas Neue', sans-serif;\n    --font-body: 'DM Sans', sans-serif;\n    --heading-spacing: 0.04em;\n    --heading-weight: 400;\n    --spec-header-spacing: 0.08em;\n    --quickfacts-shadow: 0 10px 24px rgba(15,23,42,0.06);\n  }\n  .fbp5-wrap,.fbp5-wrap * { box-sizing:border-box; margin:0; padding:0; }\n  .fbp5-wrap { font-family:var(--font-body); color:var(--text); line-height:1.7; max-width:860px; margin:0 auto; padding:0 24px 80px; display:flex; flex-direction:column; gap:52px; }\n  .fbp5-wrap[data-rh-site-profile=\"klav\"] { padding:24px 24px 80px; background:var(--page-bg); border-radius:24px; }\n  .fbp5-intro { padding-top:36px; border-top:3px solid var(--red); }\n  .fbp5-intro p { font-size:17px; color:var(--copy); line-height:1.85; margin-bottom:18px; }\n  .fbp5-intro strong { color:var(--text); font-weight:700; }\n  .rh-phonebox { background:var(--card); border:2.5px solid var(--red); border-radius:16px; overflow:hidden; margin:4px 0 0; }\n  .rh-phonebox-topbar,.rh-phonebox-bottombar { height:4px; background:linear-gradient(90deg,var(--red),var(--red-mid),var(--red)); }\n  .rh-phonebox-header { padding:24px 28px; background:var(--surface-soft); border-bottom:1px solid var(--border); display:flex; align-items:flex-start; gap:24px; }\n  .rh-phonebox-hero { width:200px; height:267px; flex-shrink:0; overflow:hidden; border-radius:12px; background:var(--surface-contrast); display:flex; align-items:center; justify-content:center; }\n  .rh-phonebox-hero img { width:100%; height:100%; object-fit:cover; display:block; }\n  .rh-phonebox-header-main { flex:1; display:flex; flex-direction:column; gap:12px; min-width:0; }\n  .rh-phonebox-tag { display:inline-flex; align-items:center; gap:8px; width:fit-content; font-size:11px; font-weight:700; letter-spacing:0.1em; text-transform:uppercase; color:var(--red); background:var(--accent-soft-bg); border:1px solid var(--accent-soft-border); border-radius:100px; padding:4px 12px; }\n  .rh-phonebox-tag-icon, .rh-phonebox-hl-icon, .rh-phonebox-section-icon { font-family:\"Apple Color Emoji\",\"Segoe UI Emoji\",\"Noto Color Emoji\",\"Segoe UI Symbol\",sans-serif; }\n  .rh-phonebox-tag-icon { font-size:12px; line-height:1; }\n  .rh-phonebox-title { font-family:var(--font-head); font-size:clamp(24px,3.5vw,38px); line-height:1; color:var(--text); letter-spacing:0.02em; margin:0; }\n  .rh-phonebox-title span { color:var(--red); }\n  .rh-phonebox-title em { font-style:normal; font-size:0.6em; color:var(--soft-label); font-family:var(--font-body); font-weight:300; letter-spacing:0.01em; }\n  .rh-phonebox-sub { font-size:14px; color:var(--muted-label); font-weight:300; margin:0; }\n  .rh-phonebox-highlights { display:grid; grid-template-columns:repeat(4,minmax(0,1fr)); gap:10px; }\n  .rh-phonebox-hl { background:#fff; border:1.5px solid var(--border); border-radius:10px; padding:12px 14px; display:flex; flex-direction:column; gap:6px; }\n  .rh-phonebox-hl-top { display:flex; align-items:flex-start; gap:8px; }\n  .rh-phonebox-hl-icon { font-size:15px; line-height:1; }\n  .rh-phonebox-hl-value { font-family:var(--font-head); font-size:22px; line-height:1; color:var(--text); letter-spacing:0.02em; }\n  .rh-phonebox-hl-value span { font-family:var(--font-body); font-size:12px; font-weight:600; color:var(--red); }\n  .rh-phonebox-hl-label { font-size:11px; line-height:1.3; color:var(--soft-label); }\n  .rh-phonebox-colors { padding:12px 28px; background:var(--surface-soft); border-bottom:1px solid var(--border); display:flex; align-items:center; gap:8px; flex-wrap:wrap; }\n  .rh-phonebox-colors-label { font-size:11px; font-weight:700; letter-spacing:0.08em; text-transform:uppercase; color:var(--soft-label); }\n  .rh-phonebox-color-chip { display:inline-flex; align-items:center; gap:6px; background:#fff; border:1.5px solid var(--border); border-radius:100px; padding:4px 12px 4px 6px; font-size:13px; color:var(--text); }\n  .rh-phonebox-color-dot { width:12px; height:12px; border-radius:50%; border:1px solid rgba(0,0,0,0.1); display:inline-block; flex-shrink:0; }\n  .rh-phonebox-section { border-bottom:1px solid var(--line-soft); display:grid; grid-template-columns:180px minmax(0,1fr); }\n  .rh-phonebox-section:last-of-type { border-bottom:none; }\n  .rh-phonebox-section-label { padding:18px 20px 18px 28px; background:var(--surface-soft); border-right:1px solid var(--border); display:flex; align-items:flex-start; gap:8px; }\n  .rh-phonebox-section-icon { font-size:16px; line-height:1; flex-shrink:0; margin-top:1px; }\n  .rh-phonebox-section-name { font-size:12px; font-weight:700; letter-spacing:0.05em; text-transform:uppercase; color:var(--muted-label); line-height:1.4; }\n  .rh-phonebox-section-values { padding:18px 28px 18px 22px; }\n  .rh-phonebox-row { margin-bottom:12px; }\n  .rh-phonebox-row:last-child { margin-bottom:0; }\n  .rh-phonebox-row-key { font-size:12px; font-weight:700; letter-spacing:0.06em; text-transform:uppercase; color:var(--red); display:block; margin-bottom:3px; }\n  .rh-phonebox-row-val { font-size:15px; line-height:1.6; color:var(--copy); }\n  .rh-phonebox-row-val strong { font-weight:600; color:var(--text); }\n  .rh-affbox { margin-top:-20px; }\n  .rh-affbox-wrap { width:100%; background:#fff; border:2px solid var(--border); border-radius:16px; overflow:hidden; }\n  .rh-affbox-grid { display:grid; grid-template-columns:repeat(auto-fit,minmax(260px,1fr)); }\n  .rh-affbox-col { padding:22px 24px; }\n  .rh-affbox-col:not(:last-child) { border-right:1px solid var(--line-soft); }\n  .rh-affbox-model { font-family:var(--font-head); font-size:22px; color:var(--text); letter-spacing:0.02em; margin-bottom:4px; line-height:1; }\n  .rh-affbox-model span { color:var(--red); }\n  .rh-affbox-from { font-size:11px; font-weight:700; letter-spacing:0.08em; text-transform:uppercase; color:var(--soft-label); margin-bottom:2px; }\n  .rh-affbox-amount { font-family:var(--font-head); font-size:36px; color:var(--red); letter-spacing:0.02em; line-height:1; margin-bottom:14px; }\n  .rh-affbox-ctas { display:flex; flex-direction:column; gap:8px; margin-bottom:14px; }\n  .rh-affbox-cta { display:inline-flex; align-items:center; justify-content:center; gap:8px; background:var(--red); color:#fff !important; text-decoration:none !important; border-radius:100px; padding:11px 22px; font-size:14px; font-weight:600; white-space:nowrap; }\n  .rh-affbox-cta.secondary { background:var(--text); }\n  .rh-affbox-cta svg { width:13px; height:13px; flex-shrink:0; }\n  .rh-affbox-coupon { display:flex; align-items:center; justify-content:space-between; gap:12px; border:1px dashed var(--accent-soft-border); background:var(--accent-soft-bg); border-radius:12px; padding:10px 12px; margin:0 0 12px; }\n  .rh-affbox-coupon-label { font-size:11px; font-weight:700; letter-spacing:0.08em; text-transform:uppercase; color:var(--red); }\n  .rh-affbox-coupon-code { font-family:var(--font-head); font-size:22px; line-height:1; color:var(--text); letter-spacing:0.03em; }\n  .rh-affbox-promo { font-size:12px; color:var(--text-muted); line-height:1.5; border-top:1px solid var(--line-soft); padding-top:12px; margin-top:2px; }\n  .rh-affbox-promo strong { color:var(--red); font-weight:600; }\n  .rh-affbox-disclaimer { margin-top:10px; font-size:12px; line-height:1.55; color:var(--text-muted); }\n  .rh-downloadbox { margin:4px 0 0; }\n  .rh-downloadbox-wrap { background:linear-gradient(135deg,var(--accent-soft-bg),#fff); border:1px solid var(--accent-soft-border); border-radius:18px; padding:20px; box-shadow:var(--quickfacts-shadow); }\n  .rh-downloadbox-head { display:flex; justify-content:space-between; align-items:flex-start; gap:16px; margin-bottom:14px; }\n  .rh-downloadbox-kicker { font-size:11px; font-weight:800; letter-spacing:0.12em; text-transform:uppercase; color:var(--red); margin-bottom:6px; }\n  .rh-downloadbox-title { font-family:var(--font-head); font-size:clamp(24px,3.4vw,34px); line-height:1; color:var(--text); letter-spacing:var(--heading-spacing); margin:0; }\n  .rh-downloadbox-count { flex-shrink:0; border:1px solid var(--accent-soft-border); background:#fff; color:var(--red); border-radius:999px; padding:6px 10px; font-size:11px; font-weight:800; text-transform:uppercase; }\n  .rh-downloadbox-links { display:flex; flex-wrap:wrap; gap:10px; }\n  .rh-downloadbox-link { display:inline-flex; align-items:center; justify-content:center; gap:8px; background:var(--red); color:#fff !important; text-decoration:none !important; border-radius:999px; padding:11px 16px; font-size:14px; font-weight:700; }\n  .rh-downloadbox-link svg { width:13px; height:13px; flex-shrink:0; }\n  .rh-downloadbox-note { margin:14px 0 0; font-size:12px; line-height:1.55; color:var(--text-muted); }\n  .rh-offerbox { margin:4px 0 0; }\n  .rh-offerbox-wrap { background:linear-gradient(135deg,var(--accent-soft-bg),#fff); border:1px solid var(--accent-soft-border); border-radius:18px; padding:20px; box-shadow:var(--quickfacts-shadow); }\n  .rh-offerbox-head { display:flex; justify-content:space-between; align-items:flex-start; gap:16px; margin-bottom:14px; }\n  .rh-offerbox-kicker { font-size:11px; font-weight:800; letter-spacing:0.12em; text-transform:uppercase; color:var(--red); margin-bottom:6px; }\n  .rh-offerbox-title { font-family:var(--font-head); font-size:clamp(24px,3.4vw,34px); line-height:1; color:var(--text); letter-spacing:var(--heading-spacing); margin:0; }\n  .rh-offerbox-count { flex-shrink:0; border:1px solid var(--accent-soft-border); background:#fff; color:var(--red); border-radius:999px; padding:6px 10px; font-size:11px; font-weight:800; text-transform:uppercase; }\n  .rh-offerbox-links { display:grid; grid-template-columns:repeat(auto-fit,minmax(220px,1fr)); gap:10px; }\n  .rh-offerbox-link { display:flex; align-items:center; justify-content:center; gap:8px; background:var(--red); color:#fff !important; text-decoration:none !important; border-radius:999px; padding:11px 16px; font-size:14px; font-weight:800; text-align:center; }\n  .rh-offerbox-link small { opacity:.82; font-size:11px; font-weight:700; }\n  .rh-offerbox-link svg { width:13px; height:13px; flex-shrink:0; }\n  .rh-offerbox-note { margin:14px 0 0; font-size:12px; line-height:1.55; color:var(--text-muted); }\n  .rh-offer-inline { margin:0; }\n  .rh-offer-inline-wrap { display:flex; align-items:center; justify-content:space-between; gap:16px; background:var(--surface-soft); border:1px solid var(--accent-soft-border); border-left:4px solid var(--red); border-radius:16px; padding:16px 18px; box-shadow:var(--quickfacts-shadow); }\n  .rh-offer-inline-kicker { font-size:10px; font-weight:800; letter-spacing:0.12em; text-transform:uppercase; color:var(--red); margin-bottom:4px; }\n  .rh-offer-inline-title { font-size:15px; line-height:1.35; color:var(--text); font-weight:800; }\n  .rh-offer-inline-store { margin-top:3px; font-size:12px; color:var(--text-muted); font-weight:700; }\n  .rh-offer-inline-cta { display:inline-flex; align-items:center; justify-content:center; gap:8px; flex-shrink:0; background:var(--red); color:#fff !important; text-decoration:none !important; border-radius:999px; padding:11px 16px; font-size:13px; font-weight:800; white-space:nowrap; }\n  .rh-offer-inline-cta svg { width:13px; height:13px; flex-shrink:0; }\n  .rh-offer-inline-note { margin:9px 0 0; font-size:12px; line-height:1.55; color:var(--text-muted); }\n  .rh-external-embed { margin:4px 0 0; }\n  .rh-external-embed-wrap { width:100%; background:#fff; border:1px solid var(--border); border-radius:16px; overflow:hidden; box-shadow:var(--quickfacts-shadow); }\n  .rh-external-embed-head { display:flex; align-items:center; justify-content:space-between; gap:12px; padding:14px 18px; background:var(--surface-soft); border-bottom:1px solid var(--border); }\n  .rh-external-embed-kicker { font-size:11px; font-weight:700; letter-spacing:0.08em; text-transform:uppercase; color:var(--red); }\n  .rh-external-embed-source { font-size:12px; color:var(--soft-label); }\n  .rh-external-embed-frame { position:relative; aspect-ratio:16 \/ 9; overflow:hidden; background:#000; min-height:0; }\n  .rh-external-embed-frame iframe, .rh-external-embed-frame > iframe { position:absolute !important; inset:0; width:100% !important; height:100% !important; min-height:0 !important; max-height:none !important; max-width:none !important; border:0; display:block; }\n  .rh-external-embed-body { padding:18px 20px; display:grid; gap:12px; }\n  .rh-external-embed-title { font-family:var(--font-head); font-size:24px; line-height:1; color:var(--text); letter-spacing:0.02em; }\n  .rh-external-embed-copy { font-size:14px; color:var(--text-muted); line-height:1.6; }\n  .rh-external-embed-link { display:inline-flex; align-items:center; gap:8px; width:fit-content; color:var(--red) !important; text-decoration:none !important; font-weight:700; }\n  .rh-external-embed-link svg { width:13px; height:13px; flex-shrink:0; }\n  .fbp5-section { display:flex; flex-direction:column; gap:0; }\n  .fbp5-section-label { display:inline-flex; align-items:center; gap:7px; font-size:11px; font-weight:700; letter-spacing:0.14em; text-transform:uppercase; color:var(--red); margin-bottom:10px; }\n  .fbp5-section-label::before { content:''; display:block; width:18px; height:2px; background:var(--red); }\n  .fbp5-section h2 { font-family:var(--font-head); font-size:clamp(27px,4vw,39px); letter-spacing:var(--heading-spacing); font-weight:var(--heading-weight); color:var(--text); line-height:1.05; margin-bottom:20px; }\n  .fbp5-section h2 span { color:var(--red); }\n  .fbp5-section p { font-size:17px; color:var(--copy); line-height:1.78; margin-bottom:14px; }\n  .fbp5-bullet-list { margin:18px 0 6px 20px; padding:0; display:flex; flex-direction:column; gap:10px; }\n  .fbp5-bullet-list li { font-size:16px; line-height:1.7; color:#1f2937; }\n  .fbp5-bullet-list li::marker { color:var(--red); }\n  .rh-quickfacts-grid { margin:18px 0 0; display:grid; grid-template-columns:repeat(auto-fit,minmax(220px,1fr)); gap:12px; }\n  .rh-quickfacts-card { background:#fff; border:1px solid var(--border); border-radius:16px; padding:14px 16px; box-shadow:var(--quickfacts-shadow); display:flex; flex-direction:column; gap:6px; }\n  .rh-quickfacts-label { font-size:12px; font-weight:800; letter-spacing:0.08em; text-transform:uppercase; color:var(--red); }\n  .rh-quickfacts-value { font-size:15px; line-height:1.6; color:#1f2937; }\n  .fbp5-highlight { background:var(--red-light); border-left:4px solid var(--red); border-radius:0 10px 10px 0; padding:15px 19px; margin:18px 0 0; font-size:16px; color:var(--copy); line-height:1.65; }\n  .fbp5-info { background:var(--surface-soft); border:1px solid var(--border); border-radius:10px; padding:15px 19px; margin:18px 0 0; font-size:15px; color:var(--text-muted); line-height:1.65; }\n  .fbp5-scheda { background:var(--card); border:1px solid var(--border); border-radius:14px; overflow:hidden; margin:20px 0 0; }\n  .fbp5-scheda-header { background:var(--red); color:#fff; padding:14px 20px; font-family:var(--font-head); font-size:20px; font-weight:var(--heading-weight); letter-spacing:var(--spec-header-spacing); }\n  .fbp5-scheda-row { display:flex; border-bottom:1px solid var(--border); }\n  .fbp5-scheda-row:last-child { border-bottom:none; }\n  .fbp5-scheda-key { width:42%; padding:11px 16px; font-size:13px; font-weight:700; color:var(--text-muted); text-transform:uppercase; letter-spacing:0.06em; background:var(--surface-soft); flex-shrink:0; }\n  .fbp5-scheda-val { padding:11px 16px; font-size:14px; color:var(--text); font-weight:500; line-height:1.5; }\n  @media(max-width:600px) {\n    .fbp5-wrap { padding:0 14px 60px; gap:40px; }\n    .fbp5-section h2 { font-size:32px; }\n    .fbp5-scheda-row { flex-direction:column; }\n    .fbp5-scheda-key { width:100%; }\n    .rh-phonebox-header { padding:18px 16px; flex-direction:column; align-items:center; }\n    .rh-phonebox-hero { width:150px; height:200px; }\n    .rh-phonebox-highlights { grid-template-columns:repeat(2,minmax(0,1fr)); }\n    .rh-phonebox-colors { padding:16px; }\n    .rh-phonebox-section { grid-template-columns:1fr; }\n    .rh-phonebox-section-label { padding:14px 20px; border-right:none; border-bottom:1px solid var(--border); }\n    .rh-phonebox-section-values { padding:14px 20px; }\n    .rh-affbox-grid { grid-template-columns:1fr; }\n    .rh-affbox-col:not(:last-child) { border-right:none; border-bottom:1px solid var(--line-soft); }\n    .rh-affbox-cta { width:100%; }\n    .rh-downloadbox-head { flex-direction:column; }\n    .rh-downloadbox-link { width:100%; }\n    .rh-offerbox-head { flex-direction:column; }\n    .rh-offerbox-links { grid-template-columns:1fr; }\n    .rh-offer-inline-wrap { flex-direction:column; align-items:flex-start; }\n    .rh-offer-inline-cta { width:100%; }\n  }\n<\/style>\n<div class=\"fbp5-wrap rh-json-article\" id=\"rh-article-mqwefhgy-9ob7h\" data-rh-article-type=\"news\" data-rh-site-profile=\"china-smart-buy\">\n  <div class=\"fbp5-intro\">\n    \n    <p>DeepSeek e l&#8217;Universita di Pechino hanno presentato DSpark, un framework pensato per ridurre il collo di bottiglia della <strong>inferenza dei modelli linguistici di grandi dimensioni<\/strong> quando le richieste simultanee aumentano. L&#8217;obiettivo e semplice da capire: far rispondere piu in fretta i modelli senza sacrificare la qualita dell&#8217;output.<\/p><p>Il dato piu interessante, per chi guarda alle prestazioni concrete, e che DSpark e gia integrato nelle versioni di anteprima <strong>DeepSeek-V4-Flash<\/strong> e <strong>DeepSeek-V4-Pro<\/strong>.<\/p>\n  <\/div>\n\n      <section class=\"fbp5-section\">\n        <div class=\"fbp5-section-label\">Tecnologia<\/div>\n        <h2>DSpark prova a risolvere il limite piu costoso dei modelli AI in produzione<\/h2>\n        <p>Il problema che DSpark prova a colpire e noto: nei modelli autoregressivi ogni nuovo token richiede un passaggio completo nella rete, quindi la latenza cresce con la lunghezza del testo generato. E uno dei motivi principali per cui i sistemi di chat AI diventano piu lenti proprio quando la richiesta e piu impegnativa.<\/p><p>Il framework parte dalla logica della <strong>speculative decoding<\/strong>, cioe la generazione rapida di candidati da parte di un modello piu leggero e la loro verifica in blocco da parte del modello principale. DSpark aggiunge pero due elementi chiave: una generazione candidata semi-autoregressiva e una verifica guidata dalla confidenza, pensate per migliorare sia il tasso di accettazione sia l&#8217;uso delle risorse di calcolo.<\/p>\n        <ul class=\"fbp5-bullet-list\"><li>La generazione candidata non e solo parallela: usa una struttura ibrida che combina un tronco principale parallelo e un modulo sequenziale leggero.<\/li><li>Il sistema assegna a ogni posizione un punteggio di confidenza per stimare la probabilita che il token venga mantenuto dopo la verifica.<\/li><li>La scelta della lunghezza da verificare non e fissa, ma adattata al carico e alla probabilita di sopravvivenza dei token.<\/li><li>DSpark e gia stato impiegato nelle preview service di DeepSeek-V4-Flash e DeepSeek-V4-Pro.<\/li><\/ul>\n        <div class=\"rh-quickfacts-grid\">\n      <div class=\"rh-quickfacts-card\">\n        <div class=\"rh-quickfacts-label\">Aumento velocita<\/div>\n        <div class=\"rh-quickfacts-value\">dal 60% all&#8217;85%<\/div>\n      <\/div>\n    \n      <div class=\"rh-quickfacts-card\">\n        <div class=\"rh-quickfacts-label\">Integrazione<\/div>\n        <div class=\"rh-quickfacts-value\">DeepSeek-V4-Flash e DeepSeek-V4-Pro preview<\/div>\n      <\/div>\n    \n      <div class=\"rh-quickfacts-card\">\n        <div class=\"rh-quickfacts-label\">Rilascio codice<\/div>\n        <div class=\"rh-quickfacts-value\">pubblicato su GitHub<\/div>\n      <\/div>\n    <\/div>\n        <div class=\"fbp5-highlight\">DSpark non punta solo a generare piu veloce, ma a spendere meglio il calcolo disponibile quando molte richieste arrivano insieme. Il punto forte e la gestione dinamica del budget di verifica, che sposta risorse dove la probabilita di accettazione e piu alta.<\/div>\n        <div class=\"fbp5-info\">L&#8217;approccio resta legato alla speculative decoding, quindi la qualita viene preservata tramite verifica e non tramite scorciatoie nella generazione finale.<\/div>\n        \n      <\/section>\n    \n\n      <section class=\"fbp5-section\">\n        <div class=\"fbp5-section-label\">Architettura<\/div>\n        <h2>La novita vera e nel mix tra parallelismo e <span>dipendenze di contesto<\/span><\/h2>\n        <p>Nel confronto con i draft model piu diffusi, DSpark prova a superare due limiti opposti. Da un lato ci sono i modelli autoregressivi come Eagle3, che mantengono buone dipendenze tra token ma diventano lenti quando il blocco candidato cresce. Dall&#8217;altro ci sono i modelli completamente paralleli come DFlash, rapidi in generazione ma piu fragili quando le posizioni successive iniziano a confliggere tra loro.<\/p><p>DSpark introduce un compromesso piu pratico: un backbone parallelo, basato su una versione migliorata di DFlash, produce gli stati nascosti e le logit di base; un modulo sequenziale leggero inserisce poi informazione di prefisso token per token. Questo consente di mantenere il vantaggio della generazione parallela iniziale, senza perdere del tutto la coerenza che aiuta l&#8217;accettazione dei token successivi.<\/p><p>Il lavoro sperimentale segnala che una versione DSpark con due layer Transformer supera, nei domini testati, l&#8217;accettazione di una versione DFlash con cinque layer. Tradotto in termini semplici: una piccola dose di dipendenza autoregressiva vale piu di un semplice aumento di profondita nella parte parallela.<\/p>\n        <ul class=\"fbp5-bullet-list\"><li>Modulo sequenziale disponibile in due varianti: testa di Markov o testa RNN.<\/li><li>Il blocco candidato massimo usato in produzione e pari a 5 token.<\/li><li>Nel deployment reale e stata scelta la testa di Markov come soluzione sequenziale.<\/li><li>Il backbone parallelo include tre layer MoE e attenzione a finestra mobile.<\/li><\/ul>\n        <div class=\"rh-quickfacts-grid\">\n      <div class=\"rh-quickfacts-card\">\n        <div class=\"rh-quickfacts-label\">Alternative confrontate<\/div>\n        <div class=\"rh-quickfacts-value\">Eagle3 e DFlash<\/div>\n      <\/div>\n    \n      <div class=\"rh-quickfacts-card\">\n        <div class=\"rh-quickfacts-label\">Versione di produzione<\/div>\n        <div class=\"rh-quickfacts-value\">DSpark-5<\/div>\n      <\/div>\n    <\/div>\n        <div class=\"fbp5-highlight\">L&#8217;idea chiave e che un po&#8217; di dipendenza sequenziale, se inserita nel punto giusto, migliora l&#8217;efficienza piu di una semplice scalata di parametri nel modello parallelo.<\/div>\n        <div class=\"fbp5-info\">Le varianti citate nel progetto puntano a ottimizzare il rapporto tra capacita di accettazione e costo di generazione, non a sostituire il modello principale.<\/div>\n        \n      <\/section>\n    \n\n      <section class=\"fbp5-section\">\n        <div class=\"fbp5-section-label\">Impatto pratico<\/div>\n        <h2>Nei test di carico DSpark migliora anche il <span>throughput complessivo<\/span><\/h2>\n        <p>Il risultato piu utile per leggere la portata del progetto e che DSpark non migliora solo la latenza percepita dal singolo utente, ma anche il rendimento globale del sistema quando i carichi crescono. Nei test online con traffico reale, il framework e stato confrontato con il baseline a singolo token MTP-1 e ha mostrato vantaggi consistenti in diversi scenari di SLA.<\/p><p>Su V4-Flash, mantenendo una velocita minima di 80 token al secondo per singolo utente, il throughput aggregato sale del <strong>51%<\/strong>. Quando l&#8217;SLA si stringe a 120 token al secondo, il baseline entra quasi nel suo limite operativo e DSpark arriva a un vantaggio di throughput dichiarato del <strong>661%<\/strong>, grazie alla capacita di mantenere un batching utile anche in condizioni piu dure.<\/p><p>Su V4-Pro, invece, il guadagno di throughput e del <strong>52%<\/strong> con SLA da 35 token al secondo e del <strong>406%<\/strong> con SLA da 50 token al secondo. Nei casi in cui il throughput complessivo e allineato, la velocita di generazione per singolo utente cresce comunque del <strong>57%-85%<\/strong>.<\/p>\n        <ul class=\"fbp5-bullet-list\"><li>Con bassa concorrenza, il scheduler assegna spesso 4-6 token di verifica per sfruttare la capacita inutilizzata.<\/li><li>Con l&#8217;aumento del carico, la lunghezza di verifica si riduce in modo graduale per limitare la contesa delle risorse.<\/li><li>Il sistema prova a massimizzare il throughput globale invece di proteggere solo il singolo campione di richiesta.<\/li><li>Un limite resta: i candidati iniziali devono comunque essere generati per intero, anche se poi parte dei token viene scartata.<\/li><\/ul>\n        <div class=\"rh-quickfacts-grid\">\n      <div class=\"rh-quickfacts-card\">\n        <div class=\"rh-quickfacts-label\">V4-Flash, SLA 80 token\/s<\/div>\n        <div class=\"rh-quickfacts-value\">+51% throughput<\/div>\n      <\/div>\n    <\/div>\n        <div class=\"fbp5-highlight\">Nel confronto operativo, DSpark e interessante perche prova a tenere insieme due obiettivi spesso in conflitto: rispondere piu velocemente e reggere meglio il traffico concorrente.<\/div>\n        <div class=\"fbp5-info\">Il progetto mostra bene dove si sta spostando l&#8217;ottimizzazione AI oggi: non solo modelli piu grandi, ma anche strategie piu intelligenti di verifica, scheduling e utilizzo della GPU.<\/div>\n        \n      <div class=\"fbp5-scheda\">\n        <div class=\"fbp5-scheda-header\">Dati di prestazione principali<\/div>\n        \n          <div class=\"fbp5-scheda-row\">\n            <div class=\"fbp5-scheda-key\">Velocita singolo utente<\/div>\n            <div class=\"fbp5-scheda-val\">+60% \/ +85%<\/div>\n          <\/div>\n        \n          <div class=\"fbp5-scheda-row\">\n            <div class=\"fbp5-scheda-key\">Throughput V4-Flash<\/div>\n            <div class=\"fbp5-scheda-val\">+51% a 80 token\/s<\/div>\n          <\/div>\n        \n          <div class=\"fbp5-scheda-row\">\n            <div class=\"fbp5-scheda-key\">Throughput V4-Flash<\/div>\n            <div class=\"fbp5-scheda-val\">+661% a 120 token\/s<\/div>\n          <\/div>\n        \n          <div class=\"fbp5-scheda-row\">\n            <div class=\"fbp5-scheda-key\">Throughput V4-Pro<\/div>\n            <div class=\"fbp5-scheda-val\">+52% a 35 token\/s<\/div>\n          <\/div>\n        \n          <div class=\"fbp5-scheda-row\">\n            <div class=\"fbp5-scheda-key\">Throughput V4-Pro<\/div>\n            <div class=\"fbp5-scheda-val\">+406% a 50 token\/s<\/div>\n          <\/div>\n        \n      <\/div>\n    \n      <\/section>\n    \n<\/div><div class=\"rh-attribution-pills\" style=\"display:flex;gap:10px;flex-wrap:wrap;justify-content:flex-start;margin:16px 0 0\"><a href=\"https:\/\/www.ithome.com\/0\/969\/379.htm\" target=\"_blank\" rel=\"noopener\" title=\"Via: VIA\" aria-label=\"Via: VIA\" style=\"display:inline-flex;align-items:center;gap:8px;padding:10px 16px;border-radius:16px;background:linear-gradient(135deg,#fff7ed,#ffedd5);color:#c2410c;text-decoration:none;font-size:11px;font-weight:900;letter-spacing:.08em;text-transform:uppercase;border:1px solid rgba(249,115,22,.20);box-shadow:0 8px 22px rgba(249,115,22,.08)\">Via<\/a><\/div>\n","protected":false},"excerpt":{"rendered":"<p>DeepSeek e Universita di Pechino presentano DSpark, un framework open source che accelera l&#8217;inferenza dei modelli AI e migliora il throughput in produzione.<\/p>\n","protected":false},"author":9,"featured_media":33193,"comment_status":"open","ping_status":"open","sticky":false,"template":"","format":"standard","meta":{"_exactmetrics_skip_tracking":false,"_yoast_wpseo_focuskw_text_input":"DSpark DeepSeek inferenza AI","footnotes":""},"categories":[480],"tags":[933],"class_list":["post-33194","post","type-post","status-publish","format-standard","has-post-thumbnail","hentry","category-news","tag-ai"],"blocksy_meta":{"styles_descriptor":{"styles":{"desktop":"","tablet":"","mobile":""},"google_fonts":[],"version":8}},"yoast_head":"<!-- This site is optimized with the Yoast SEO plugin v28.6 - https:\/\/yoast.com\/product\/yoast-seo-wordpress\/ -->\n<title>DSpark di DeepSeek accelera l&#039;inferenza AI fino all&#039;85%<\/title>\n<meta name=\"description\" content=\"DeepSeek e Universita di Pechino lanciano DSpark: inferenza piu rapida, throughput migliore e codice gia su GitHub.\" \/>\n<meta name=\"robots\" content=\"index, follow, max-snippet:-1, max-image-preview:large, max-video-preview:-1\" \/>\n<link rel=\"canonical\" href=\"https:\/\/www.chinasmartbuy.com\/blog\/dspark-deepseek-inferenza-ai\/\" \/>\n<meta property=\"og:locale\" content=\"it_IT\" \/>\n<meta property=\"og:type\" content=\"article\" \/>\n<meta property=\"og:title\" content=\"DSpark di DeepSeek accelera l&#039;inferenza AI fino all&#039;85%\" \/>\n<meta property=\"og:description\" content=\"DeepSeek e Universita di Pechino lanciano DSpark: inferenza piu rapida, throughput migliore e codice gia su GitHub.\" \/>\n<meta property=\"og:url\" content=\"https:\/\/www.chinasmartbuy.com\/blog\/dspark-deepseek-inferenza-ai\/\" \/>\n<meta property=\"og:site_name\" content=\"ChinaSmartBuy\" \/>\n<meta property=\"article:publisher\" content=\"http:\/\/facebook.com\/chinasmartbuy\" \/>\n<meta property=\"article:published_time\" content=\"2026-06-27T15:05:00+00:00\" \/>\n<meta property=\"og:image\" content=\"https:\/\/www.chinasmartbuy.com\/blog\/wp-content\/uploads\/2026\/06\/dspark-di-deepseek-accelera-l-inferenza-ai-fino-all-85-68dd2c21-582a-42fb-95e4-c87b-5uvfp3zov9.jpg\" \/>\n\t<meta property=\"og:image:width\" content=\"820\" \/>\n\t<meta property=\"og:image:height\" content=\"453\" \/>\n\t<meta property=\"og:image:type\" content=\"image\/jpeg\" \/>\n<meta name=\"author\" content=\"Giuliano Iervolino\" \/>\n<meta name=\"twitter:card\" content=\"summary_large_image\" \/>\n<meta name=\"twitter:label1\" content=\"Scritto da\" \/>\n\t<meta name=\"twitter:data1\" content=\"Giuliano Iervolino\" \/>\n\t<meta name=\"twitter:label2\" content=\"Tempo di lettura stimato\" \/>\n\t<meta name=\"twitter:data2\" content=\"5 minuti\" \/>\n<script type=\"application\/ld+json\" class=\"yoast-schema-graph\">{\"@context\":\"https:\\\/\\\/schema.org\",\"@graph\":[{\"@type\":\"Article\",\"@id\":\"https:\\\/\\\/www.chinasmartbuy.com\\\/blog\\\/dspark-deepseek-inferenza-ai\\\/#article\",\"isPartOf\":{\"@id\":\"https:\\\/\\\/www.chinasmartbuy.com\\\/blog\\\/dspark-deepseek-inferenza-ai\\\/\"},\"author\":{\"name\":\"Giuliano Iervolino\",\"@id\":\"https:\\\/\\\/www.chinasmartbuy.com\\\/blog\\\/#\\\/schema\\\/person\\\/300686f0d3c1d4edc8811bd3da82337e\"},\"headline\":\"DSpark di DeepSeek accelera l&#8217;inferenza AI fino all&#8217;85%\",\"datePublished\":\"2026-06-27T15:05:00+00:00\",\"mainEntityOfPage\":{\"@id\":\"https:\\\/\\\/www.chinasmartbuy.com\\\/blog\\\/dspark-deepseek-inferenza-ai\\\/\"},\"wordCount\":1004,\"commentCount\":0,\"publisher\":{\"@id\":\"https:\\\/\\\/www.chinasmartbuy.com\\\/blog\\\/#organization\"},\"image\":{\"@id\":\"https:\\\/\\\/www.chinasmartbuy.com\\\/blog\\\/dspark-deepseek-inferenza-ai\\\/#primaryimage\"},\"thumbnailUrl\":\"https:\\\/\\\/www.chinasmartbuy.com\\\/blog\\\/wp-content\\\/uploads\\\/2026\\\/06\\\/dspark-di-deepseek-accelera-l-inferenza-ai-fino-all-85-68dd2c21-582a-42fb-95e4-c87b-5uvfp3zov9.jpg\",\"keywords\":[\"ai\"],\"articleSection\":[\"News\"],\"inLanguage\":\"it-IT\",\"potentialAction\":[{\"@type\":\"CommentAction\",\"name\":\"Comment\",\"target\":[\"https:\\\/\\\/www.chinasmartbuy.com\\\/blog\\\/dspark-deepseek-inferenza-ai\\\/#respond\"]}]},{\"@type\":\"WebPage\",\"@id\":\"https:\\\/\\\/www.chinasmartbuy.com\\\/blog\\\/dspark-deepseek-inferenza-ai\\\/\",\"url\":\"https:\\\/\\\/www.chinasmartbuy.com\\\/blog\\\/dspark-deepseek-inferenza-ai\\\/\",\"name\":\"DSpark di DeepSeek accelera l'inferenza AI fino all'85%\",\"isPartOf\":{\"@id\":\"https:\\\/\\\/www.chinasmartbuy.com\\\/blog\\\/#website\"},\"primaryImageOfPage\":{\"@id\":\"https:\\\/\\\/www.chinasmartbuy.com\\\/blog\\\/dspark-deepseek-inferenza-ai\\\/#primaryimage\"},\"image\":{\"@id\":\"https:\\\/\\\/www.chinasmartbuy.com\\\/blog\\\/dspark-deepseek-inferenza-ai\\\/#primaryimage\"},\"thumbnailUrl\":\"https:\\\/\\\/www.chinasmartbuy.com\\\/blog\\\/wp-content\\\/uploads\\\/2026\\\/06\\\/dspark-di-deepseek-accelera-l-inferenza-ai-fino-all-85-68dd2c21-582a-42fb-95e4-c87b-5uvfp3zov9.jpg\",\"datePublished\":\"2026-06-27T15:05:00+00:00\",\"description\":\"DeepSeek e Universita di Pechino lanciano DSpark: inferenza piu rapida, throughput migliore e codice gia su GitHub.\",\"breadcrumb\":{\"@id\":\"https:\\\/\\\/www.chinasmartbuy.com\\\/blog\\\/dspark-deepseek-inferenza-ai\\\/#breadcrumb\"},\"inLanguage\":\"it-IT\",\"potentialAction\":[{\"@type\":\"ReadAction\",\"target\":[\"https:\\\/\\\/www.chinasmartbuy.com\\\/blog\\\/dspark-deepseek-inferenza-ai\\\/\"]}]},{\"@type\":\"ImageObject\",\"inLanguage\":\"it-IT\",\"@id\":\"https:\\\/\\\/www.chinasmartbuy.com\\\/blog\\\/dspark-deepseek-inferenza-ai\\\/#primaryimage\",\"url\":\"https:\\\/\\\/www.chinasmartbuy.com\\\/blog\\\/wp-content\\\/uploads\\\/2026\\\/06\\\/dspark-di-deepseek-accelera-l-inferenza-ai-fino-all-85-68dd2c21-582a-42fb-95e4-c87b-5uvfp3zov9.jpg\",\"contentUrl\":\"https:\\\/\\\/www.chinasmartbuy.com\\\/blog\\\/wp-content\\\/uploads\\\/2026\\\/06\\\/dspark-di-deepseek-accelera-l-inferenza-ai-fino-all-85-68dd2c21-582a-42fb-95e4-c87b-5uvfp3zov9.jpg\",\"width\":820,\"height\":453,\"caption\":\"DSpark di DeepSeek accelera l'inferenza AI fino all'85%\"},{\"@type\":\"BreadcrumbList\",\"@id\":\"https:\\\/\\\/www.chinasmartbuy.com\\\/blog\\\/dspark-deepseek-inferenza-ai\\\/#breadcrumb\",\"itemListElement\":[{\"@type\":\"ListItem\",\"position\":1,\"name\":\"Home\",\"item\":\"https:\\\/\\\/www.chinasmartbuy.com\\\/blog\\\/\"},{\"@type\":\"ListItem\",\"position\":2,\"name\":\"DSpark di DeepSeek accelera l&#8217;inferenza AI fino all&#8217;85%\"}]},{\"@type\":\"WebSite\",\"@id\":\"https:\\\/\\\/www.chinasmartbuy.com\\\/blog\\\/#website\",\"url\":\"https:\\\/\\\/www.chinasmartbuy.com\\\/blog\\\/\",\"name\":\"ChinaSmartBuy\",\"description\":\"NEWS \u00b7 GUIDE \u00b7 OFFERTE\",\"publisher\":{\"@id\":\"https:\\\/\\\/www.chinasmartbuy.com\\\/blog\\\/#organization\"},\"potentialAction\":[{\"@type\":\"SearchAction\",\"target\":{\"@type\":\"EntryPoint\",\"urlTemplate\":\"https:\\\/\\\/www.chinasmartbuy.com\\\/blog\\\/?s={search_term_string}\"},\"query-input\":{\"@type\":\"PropertyValueSpecification\",\"valueRequired\":true,\"valueName\":\"search_term_string\"}}],\"inLanguage\":\"it-IT\"},{\"@type\":\"Organization\",\"@id\":\"https:\\\/\\\/www.chinasmartbuy.com\\\/blog\\\/#organization\",\"name\":\"ChinaSmartBuy\",\"url\":\"https:\\\/\\\/www.chinasmartbuy.com\\\/blog\\\/\",\"logo\":{\"@type\":\"ImageObject\",\"inLanguage\":\"it-IT\",\"@id\":\"https:\\\/\\\/www.chinasmartbuy.com\\\/blog\\\/#\\\/schema\\\/logo\\\/image\\\/\",\"url\":\"https:\\\/\\\/www.chinasmartbuy.com\\\/blog\\\/wp-content\\\/uploads\\\/2019\\\/06\\\/logo-chinasmartbuy-e1560864925407.png\",\"contentUrl\":\"https:\\\/\\\/www.chinasmartbuy.com\\\/blog\\\/wp-content\\\/uploads\\\/2019\\\/06\\\/logo-chinasmartbuy-e1560864925407.png\",\"width\":100,\"height\":100,\"caption\":\"ChinaSmartBuy\"},\"image\":{\"@id\":\"https:\\\/\\\/www.chinasmartbuy.com\\\/blog\\\/#\\\/schema\\\/logo\\\/image\\\/\"},\"sameAs\":[\"http:\\\/\\\/facebook.com\\\/chinasmartbuy\",\"https:\\\/\\\/t.me\\\/chinasmartbuy\"]},{\"@type\":\"Person\",\"@id\":\"https:\\\/\\\/www.chinasmartbuy.com\\\/blog\\\/#\\\/schema\\\/person\\\/300686f0d3c1d4edc8811bd3da82337e\",\"name\":\"Giuliano Iervolino\",\"image\":{\"@type\":\"ImageObject\",\"inLanguage\":\"it-IT\",\"@id\":\"https:\\\/\\\/www.chinasmartbuy.com\\\/blog\\\/wp-content\\\/uploads\\\/2026\\\/09\\\/Giuliano-Iervolino-96x96.png\",\"url\":\"https:\\\/\\\/www.chinasmartbuy.com\\\/blog\\\/wp-content\\\/uploads\\\/2026\\\/09\\\/Giuliano-Iervolino-96x96.png\",\"contentUrl\":\"https:\\\/\\\/www.chinasmartbuy.com\\\/blog\\\/wp-content\\\/uploads\\\/2026\\\/09\\\/Giuliano-Iervolino-96x96.png\",\"caption\":\"Giuliano Iervolino\"},\"url\":\"https:\\\/\\\/www.chinasmartbuy.com\\\/blog\\\/author\\\/ju\\\/\"}]}<\/script>\n<!-- \/ Yoast SEO plugin. -->","yoast_head_json":{"title":"DSpark di DeepSeek accelera l'inferenza AI fino all'85%","description":"DeepSeek e Universita di Pechino lanciano DSpark: inferenza piu rapida, throughput migliore e codice gia su GitHub.","robots":{"index":"index","follow":"follow","max-snippet":"max-snippet:-1","max-image-preview":"max-image-preview:large","max-video-preview":"max-video-preview:-1"},"canonical":"https:\/\/www.chinasmartbuy.com\/blog\/dspark-deepseek-inferenza-ai\/","og_locale":"it_IT","og_type":"article","og_title":"DSpark di DeepSeek accelera l'inferenza AI fino all'85%","og_description":"DeepSeek e Universita di Pechino lanciano DSpark: inferenza piu rapida, throughput migliore e codice gia su GitHub.","og_url":"https:\/\/www.chinasmartbuy.com\/blog\/dspark-deepseek-inferenza-ai\/","og_site_name":"ChinaSmartBuy","article_publisher":"http:\/\/facebook.com\/chinasmartbuy","article_published_time":"2026-06-27T15:05:00+00:00","og_image":[{"width":820,"height":453,"url":"https:\/\/www.chinasmartbuy.com\/blog\/wp-content\/uploads\/2026\/06\/dspark-di-deepseek-accelera-l-inferenza-ai-fino-all-85-68dd2c21-582a-42fb-95e4-c87b-5uvfp3zov9.jpg","type":"image\/jpeg"}],"author":"Giuliano Iervolino","twitter_card":"summary_large_image","twitter_misc":{"Scritto da":"Giuliano Iervolino","Tempo di lettura stimato":"5 minuti"},"schema":{"@context":"https:\/\/schema.org","@graph":[{"@type":"Article","@id":"https:\/\/www.chinasmartbuy.com\/blog\/dspark-deepseek-inferenza-ai\/#article","isPartOf":{"@id":"https:\/\/www.chinasmartbuy.com\/blog\/dspark-deepseek-inferenza-ai\/"},"author":{"name":"Giuliano Iervolino","@id":"https:\/\/www.chinasmartbuy.com\/blog\/#\/schema\/person\/300686f0d3c1d4edc8811bd3da82337e"},"headline":"DSpark di DeepSeek accelera l&#8217;inferenza AI fino all&#8217;85%","datePublished":"2026-06-27T15:05:00+00:00","mainEntityOfPage":{"@id":"https:\/\/www.chinasmartbuy.com\/blog\/dspark-deepseek-inferenza-ai\/"},"wordCount":1004,"commentCount":0,"publisher":{"@id":"https:\/\/www.chinasmartbuy.com\/blog\/#organization"},"image":{"@id":"https:\/\/www.chinasmartbuy.com\/blog\/dspark-deepseek-inferenza-ai\/#primaryimage"},"thumbnailUrl":"https:\/\/www.chinasmartbuy.com\/blog\/wp-content\/uploads\/2026\/06\/dspark-di-deepseek-accelera-l-inferenza-ai-fino-all-85-68dd2c21-582a-42fb-95e4-c87b-5uvfp3zov9.jpg","keywords":["ai"],"articleSection":["News"],"inLanguage":"it-IT","potentialAction":[{"@type":"CommentAction","name":"Comment","target":["https:\/\/www.chinasmartbuy.com\/blog\/dspark-deepseek-inferenza-ai\/#respond"]}]},{"@type":"WebPage","@id":"https:\/\/www.chinasmartbuy.com\/blog\/dspark-deepseek-inferenza-ai\/","url":"https:\/\/www.chinasmartbuy.com\/blog\/dspark-deepseek-inferenza-ai\/","name":"DSpark di DeepSeek accelera l'inferenza AI fino all'85%","isPartOf":{"@id":"https:\/\/www.chinasmartbuy.com\/blog\/#website"},"primaryImageOfPage":{"@id":"https:\/\/www.chinasmartbuy.com\/blog\/dspark-deepseek-inferenza-ai\/#primaryimage"},"image":{"@id":"https:\/\/www.chinasmartbuy.com\/blog\/dspark-deepseek-inferenza-ai\/#primaryimage"},"thumbnailUrl":"https:\/\/www.chinasmartbuy.com\/blog\/wp-content\/uploads\/2026\/06\/dspark-di-deepseek-accelera-l-inferenza-ai-fino-all-85-68dd2c21-582a-42fb-95e4-c87b-5uvfp3zov9.jpg","datePublished":"2026-06-27T15:05:00+00:00","description":"DeepSeek e Universita di Pechino lanciano DSpark: inferenza piu rapida, throughput migliore e codice gia su GitHub.","breadcrumb":{"@id":"https:\/\/www.chinasmartbuy.com\/blog\/dspark-deepseek-inferenza-ai\/#breadcrumb"},"inLanguage":"it-IT","potentialAction":[{"@type":"ReadAction","target":["https:\/\/www.chinasmartbuy.com\/blog\/dspark-deepseek-inferenza-ai\/"]}]},{"@type":"ImageObject","inLanguage":"it-IT","@id":"https:\/\/www.chinasmartbuy.com\/blog\/dspark-deepseek-inferenza-ai\/#primaryimage","url":"https:\/\/www.chinasmartbuy.com\/blog\/wp-content\/uploads\/2026\/06\/dspark-di-deepseek-accelera-l-inferenza-ai-fino-all-85-68dd2c21-582a-42fb-95e4-c87b-5uvfp3zov9.jpg","contentUrl":"https:\/\/www.chinasmartbuy.com\/blog\/wp-content\/uploads\/2026\/06\/dspark-di-deepseek-accelera-l-inferenza-ai-fino-all-85-68dd2c21-582a-42fb-95e4-c87b-5uvfp3zov9.jpg","width":820,"height":453,"caption":"DSpark di DeepSeek accelera l'inferenza AI fino all'85%"},{"@type":"BreadcrumbList","@id":"https:\/\/www.chinasmartbuy.com\/blog\/dspark-deepseek-inferenza-ai\/#breadcrumb","itemListElement":[{"@type":"ListItem","position":1,"name":"Home","item":"https:\/\/www.chinasmartbuy.com\/blog\/"},{"@type":"ListItem","position":2,"name":"DSpark di DeepSeek accelera l&#8217;inferenza AI fino all&#8217;85%"}]},{"@type":"WebSite","@id":"https:\/\/www.chinasmartbuy.com\/blog\/#website","url":"https:\/\/www.chinasmartbuy.com\/blog\/","name":"ChinaSmartBuy","description":"NEWS \u00b7 GUIDE \u00b7 OFFERTE","publisher":{"@id":"https:\/\/www.chinasmartbuy.com\/blog\/#organization"},"potentialAction":[{"@type":"SearchAction","target":{"@type":"EntryPoint","urlTemplate":"https:\/\/www.chinasmartbuy.com\/blog\/?s={search_term_string}"},"query-input":{"@type":"PropertyValueSpecification","valueRequired":true,"valueName":"search_term_string"}}],"inLanguage":"it-IT"},{"@type":"Organization","@id":"https:\/\/www.chinasmartbuy.com\/blog\/#organization","name":"ChinaSmartBuy","url":"https:\/\/www.chinasmartbuy.com\/blog\/","logo":{"@type":"ImageObject","inLanguage":"it-IT","@id":"https:\/\/www.chinasmartbuy.com\/blog\/#\/schema\/logo\/image\/","url":"https:\/\/www.chinasmartbuy.com\/blog\/wp-content\/uploads\/2019\/06\/logo-chinasmartbuy-e1560864925407.png","contentUrl":"https:\/\/www.chinasmartbuy.com\/blog\/wp-content\/uploads\/2019\/06\/logo-chinasmartbuy-e1560864925407.png","width":100,"height":100,"caption":"ChinaSmartBuy"},"image":{"@id":"https:\/\/www.chinasmartbuy.com\/blog\/#\/schema\/logo\/image\/"},"sameAs":["http:\/\/facebook.com\/chinasmartbuy","https:\/\/t.me\/chinasmartbuy"]},{"@type":"Person","@id":"https:\/\/www.chinasmartbuy.com\/blog\/#\/schema\/person\/300686f0d3c1d4edc8811bd3da82337e","name":"Giuliano Iervolino","image":{"@type":"ImageObject","inLanguage":"it-IT","@id":"https:\/\/www.chinasmartbuy.com\/blog\/wp-content\/uploads\/2026\/09\/Giuliano-Iervolino-96x96.png","url":"https:\/\/www.chinasmartbuy.com\/blog\/wp-content\/uploads\/2026\/09\/Giuliano-Iervolino-96x96.png","contentUrl":"https:\/\/www.chinasmartbuy.com\/blog\/wp-content\/uploads\/2026\/09\/Giuliano-Iervolino-96x96.png","caption":"Giuliano Iervolino"},"url":"https:\/\/www.chinasmartbuy.com\/blog\/author\/ju\/"}]}},"_links":{"self":[{"href":"https:\/\/www.chinasmartbuy.com\/blog\/wp-json\/wp\/v2\/posts\/33194","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/www.chinasmartbuy.com\/blog\/wp-json\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/www.chinasmartbuy.com\/blog\/wp-json\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/www.chinasmartbuy.com\/blog\/wp-json\/wp\/v2\/users\/9"}],"replies":[{"embeddable":true,"href":"https:\/\/www.chinasmartbuy.com\/blog\/wp-json\/wp\/v2\/comments?post=33194"}],"version-history":[{"count":1,"href":"https:\/\/www.chinasmartbuy.com\/blog\/wp-json\/wp\/v2\/posts\/33194\/revisions"}],"predecessor-version":[{"id":33195,"href":"https:\/\/www.chinasmartbuy.com\/blog\/wp-json\/wp\/v2\/posts\/33194\/revisions\/33195"}],"wp:featuredmedia":[{"embeddable":true,"href":"https:\/\/www.chinasmartbuy.com\/blog\/wp-json\/wp\/v2\/media\/33193"}],"wp:attachment":[{"href":"https:\/\/www.chinasmartbuy.com\/blog\/wp-json\/wp\/v2\/media?parent=33194"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/www.chinasmartbuy.com\/blog\/wp-json\/wp\/v2\/categories?post=33194"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/www.chinasmartbuy.com\/blog\/wp-json\/wp\/v2\/tags?post=33194"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}