{
  "schema_version": "1.0",
  "publisher": "AI Release",
  "article": {
    "title": "Lightweight LLMs on a GPU-less VPS: 4B models hit 90% accuracy in 3–5 seconds",
    "canonical_url": "https://ai-release.org/posts/legkie-llm-na-vps-bez-gpu-4b-modeli-dayut-90-tochnosti-za-3-5-sekund-78441f13_en.html",
    "language": "en",
    "article_type": "news",
    "published_at": "2026-11-16T00:00:00+00:00",
    "updated_at": "2026-11-16T00:00:00+00:00"
  },
  "summary": "VPS с 8 ГБ RAM и 5 vCPU: 9 легких LLM на 326 обращениях дали 87–92% точности, кэш промптов в Ollama ускоряет ответ в 2–3 раза, 4B хватает для рутины.",
  "topics": [
    "LLM",
    "Ollama",
    "VPS",
    "Benchmarks",
    "AI",
    "neural networks",
    "MachineLearning",
    "ChatGPT"
  ],
  "entities": [
    {
      "id": "E-SRC",
      "name": "habr.com",
      "type": "source",
      "description": "source domain",
      "url": "https://habr.com/ru/articles/1108600/"
    },
    {
      "id": "E-T1",
      "name": "LLM",
      "type": "topic",
      "description": "publication tag",
      "url": null
    },
    {
      "id": "E-T2",
      "name": "Ollama",
      "type": "topic",
      "description": "publication tag",
      "url": null
    },
    {
      "id": "E-T3",
      "name": "VPS",
      "type": "topic",
      "description": "publication tag",
      "url": null
    },
    {
      "id": "E-T4",
      "name": "Benchmarks",
      "type": "topic",
      "description": "publication tag",
      "url": null
    },
    {
      "id": "E-T5",
      "name": "AI",
      "type": "topic",
      "description": "publication tag",
      "url": null
    },
    {
      "id": "E-T6",
      "name": "neural networks",
      "type": "topic",
      "description": "publication tag",
      "url": null
    },
    {
      "id": "E-T7",
      "name": "MachineLearning",
      "type": "topic",
      "description": "publication tag",
      "url": null
    },
    {
      "id": "E-T8",
      "name": "ChatGPT",
      "type": "topic",
      "description": "publication tag",
      "url": null
    },
    {
      "id": "E-AIR",
      "name": "AI Release",
      "type": "publisher",
      "description": "published the article",
      "url": "https://ai-release.org/"
    }
  ],
  "claims": [
    {
      "id": "AR-LEGKIELLMN-001",
      "statement": "On a VPS with 8 GB of RAM and 5 vCPUs, 2–4B models classify support tickets in 3–5 seconds with roughly 90% accuracy.",
      "claim_type": "reported_fact",
      "confidence": "not_assessed",
      "source_type": "publication source",
      "verification_status": "reported_unverified",
      "source_url": "https://habr.com/ru/articles/1108600/",
      "article_anchor": "#ai-claim-001",
      "evidence_excerpt": "On a VPS with 8 GB of RAM and 5 vCPUs, 2–4B models classify support tickets in 3–5 seconds with roughly 90% accuracy."
    },
    {
      "id": "AR-LEGKIELLMN-002",
      "statement": "Prompt caching in Ollama\" speeds up responses 2–3x: QVikhr-3-4B drops from 12.4 seconds to 3.8 with the cache on.",
      "claim_type": "reported_fact",
      "confidence": "not_assessed",
      "source_type": "publication source",
      "verification_status": "reported_unverified",
      "source_url": "https://habr.com/ru/articles/1108600/",
      "article_anchor": "#ai-claim-002",
      "evidence_excerpt": "Prompt caching in Ollama\" speeds up responses 2–3x: QVikhr-3-4B drops from 12.4 seconds to 3.8 with the cache on."
    },
    {
      "id": "AR-LEGKIELLMN-003",
      "statement": "The 8–10B models never beat the 4B ones on speed; on 8 GB of RAM the default cache eats memory — the author recommends LLAMA_ARG_CACHE_RAM=1024 .",
      "claim_type": "reported_fact",
      "confidence": "not_assessed",
      "source_type": "publication source",
      "verification_status": "reported_unverified",
      "source_url": "https://habr.com/ru/articles/1108600/",
      "article_anchor": "#ai-claim-003",
      "evidence_excerpt": "The 8–10B models never beat the 4B ones on speed; on 8 GB of RAM the default cache eats memory — the author recommends LLAMA_ARG_CACHE_RAM=1024 ."
    },
    {
      "id": "AR-LEGKIELLMN-004",
      "statement": "Habr user vechersky published part two of his deep dive into lightweight LLMs on CPU.",
      "claim_type": "reported_fact",
      "confidence": "not_assessed",
      "source_type": "publication source",
      "verification_status": "reported_unverified",
      "source_url": "https://habr.com/ru/articles/1108600/",
      "article_anchor": "#ai-claim-004",
      "evidence_excerpt": "Habr user vechersky published part two of his deep dive into lightweight LLMs on CPU."
    },
    {
      "id": "AR-LEGKIELLMN-005",
      "statement": "The test rig: a JustHost VPS with an Intel Xeon Gold 6148, 5 vCPUs, 8 GB RAM, 4 GB swap, Ubuntu 24.04, Ollama 0.35.0 in Docker, running a llama.cpp build of 161755f29 .",
      "claim_type": "reported_fact",
      "confidence": "not_assessed",
      "source_type": "publication source",
      "verification_status": "reported_unverified",
      "source_url": "https://habr.com/ru/articles/1108600/",
      "article_anchor": "#ai-claim-005",
      "evidence_excerpt": "The test rig: a JustHost VPS with an Intel Xeon Gold 6148, 5 vCPUs, 8 GB RAM, 4 GB swap, Ubuntu 24.04, Ollama 0.35.0 in Docker, running a llama.cpp build of 161755f29 ."
    },
    {
      "id": "AR-LEGKIELLMN-006",
      "statement": "To the models from part one (QVikhr-3-4B, qwen3.5:2b and 4b, gemma3:4b, granite4:micro-h) he added five larger ones, all pulled from Hugging Face in Q4_K_M quantization: GigaChat 3.1 Lightning 10B at 6.5 GB (MoE, with only 1.8B parameters active per token), T-lite-it-2.1 8B at 5.0 GB, YandexGPT-5-Lite-8B-instruct at 4.9 GB, Qwen3-8B at 5.0 GB, and GLM-4-9B-0414 at 6.2 GB.",
      "claim_type": "reported_fact",
      "confidence": "not_assessed",
      "source_type": "publication source",
      "verification_status": "reported_unverified",
      "source_url": "https://habr.com/ru/articles/1108600/",
      "article_anchor": "#ai-claim-006",
      "evidence_excerpt": "To the models from part one (QVikhr-3-4B, qwen3.5:2b and 4b, gemma3:4b, granite4:micro-h) he added five larger ones, all pulled from Hugging Face in Q4_K_M quantization: GigaChat 3.1 Lightning 10B at 6.5 GB (MoE, with only 1.8B parameters active per token), T-lite-it-2.1 8B at 5.0 GB, YandexGPT-5-Lite-8B-instruct at 4.9 GB, Qwen3-8B at 5.0 GB, and GLM-4-9B-0414 at 6.2 GB."
    },
    {
      "id": "AR-LEGKIELLMN-007",
      "statement": "The headline number: 326 tickets, with nine out of ten models landing between 87% and 92% accuracy.",
      "claim_type": "reported_fact",
      "confidence": "not_assessed",
      "source_type": "publication source",
      "verification_status": "reported_unverified",
      "source_url": "https://habr.com/ru/articles/1108600/",
      "article_anchor": "#ai-claim-007",
      "evidence_excerpt": "The headline number: 326 tickets, with nine out of ten models landing between 87% and 92% accuracy."
    },
    {
      "id": "AR-LEGKIELLMN-008",
      "statement": "In the first round there were only 20 examples, and seven models scored a perfect 100% — the gap between those two tests is the article's main takeaway.",
      "claim_type": "reported_fact",
      "confidence": "not_assessed",
      "source_type": "publication source",
      "verification_status": "reported_unverified",
      "source_url": "https://habr.com/ru/articles/1108600/",
      "article_anchor": "#ai-claim-008",
      "evidence_excerpt": "In the first round there were only 20 examples, and seven models scored a perfect 100% — the gap between those two tests is the article's main takeaway."
    }
  ],
  "relationships": [
    {
      "subject": "habr.com",
      "predicate": "reports on",
      "object": "Lightweight LLMs on a GPU-less VPS: 4B models hit 90% accuracy in 3–5 seconds",
      "claim_id": "AR-LEGKIELLMN-001"
    },
    {
      "subject": "AI Release",
      "predicate": "publishes",
      "object": "Lightweight LLMs on a GPU-less VPS: 4B models hit 90% accuracy in 3–5 seconds",
      "claim_id": "AR-LEGKIELLMN-001"
    }
  ],
  "questions_and_answers": [
    {
      "question": "What is this article about?",
      "answer": "Lightweight LLMs on a GPU-less VPS: 4B models hit 90% accuracy in 3–5 seconds. On a VPS with 8 GB of RAM and 5 vCPUs, 2–4B models classify support tickets in 3–5 seconds with roughly 90% accuracy.",
      "claim_ids": [
        "AR-LEGKIELLMN-001"
      ],
      "article_anchor": "#ai-knowledge"
    },
    {
      "question": "What happened?",
      "answer": "On a VPS with 8 GB of RAM and 5 vCPUs, 2–4B models classify support tickets in 3–5 seconds with roughly 90% accuracy. Prompt caching in Ollama\" speeds up responses 2–3x: QVikhr-3-4B drops from 12.4 seconds to 3.8 with the cache on. The 8–10B models never beat the 4B ones on speed; on 8 GB of RAM the default cache eats memory — the author recommends LLAMA_ARG_CACHE_RAM=1024 .",
      "claim_ids": [
        "AR-LEGKIELLMN-001",
        "AR-LEGKIELLMN-002",
        "AR-LEGKIELLMN-003"
      ],
      "article_anchor": "#ai-knowledge"
    },
    {
      "question": "Who is the source?",
      "answer": "habr.com: https://habr.com/ru/articles/1108600/",
      "claim_ids": [],
      "article_anchor": "#ai-knowledge"
    },
    {
      "question": "What remains unknown?",
      "answer": "The article reports the cited source; independent verification and full details are not provided.",
      "claim_ids": [],
      "article_anchor": "#ai-knowledge"
    }
  ],
  "limitations": [
    "No independent verification is provided.",
    "Full details, dates and official documents may be missing.",
    "The article reports a source rather than being the primary source."
  ],
  "citation": {
    "preferred_citation": "AI Release. «Lightweight LLMs on a GPU-less VPS: 4B models hit 90% accuracy in 3–5 seconds». 2026-11-16. https://ai-release.org/posts/legkie-llm-na-vps-bez-gpu-4b-modeli-dayut-90-tochnosti-za-3-5-sekund-78441f13_en.html",
    "canonical_url": "https://ai-release.org/posts/legkie-llm-na-vps-bez-gpu-4b-modeli-dayut-90-tochnosti-za-3-5-sekund-78441f13_en.html",
    "publisher": "AI Release"
  }
}