{
  "text": "Quen 3.8 27B. A serious local codingmodel. But what will it run on? Dense 27\nbillion parameters, 4-bit quantization,about 18 GB. But loading the weights isnot the whole story. 16 gigs, not a Q4.\nYou would need aggressive quantizationor CPU offload. 24 gigs is where it getsinteresting. Q4 fits a single 24 gig\nGPU. A 24 gig Mac [music] squeezes itin, but that memory is shared with MacOS. It fits. Not the tier I would buy.\n32 gigs of unified memory.>> [music]>> Now it makes real sense on a Mac. 48 to64, real breathing room for context and\nagents. And a 32 gig GPU keeps thentire model accelerated. So yes, aserious 27 billion parameter model runs\non consumer hardware. On a Mac, 32 gigs,64 to use it like an agent. Quen 3.8 isonly the beginning. Comment the model\nyou want me to test next.",
  "language": "en",
  "duration": 65.0,
  "segments": [
    {
      "id": 0,
      "start": 0.0,
      "end": 7.36,
      "text": "Quen 3.8 27B. A serious local codingmodel. But what will it run on? Dense 27",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 1,
      "start": 7.36,
      "end": 17.36,
      "text": "billion parameters, 4-bit quantization,about 18 GB. But loading the weights isnot the whole story. 16 gigs, not a Q4.",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 2,
      "start": 17.36,
      "end": 25.92,
      "text": "You would need aggressive quantizationor CPU offload. 24 gigs is where it getsinteresting. Q4 fits a single 24 gig",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 3,
      "start": 25.92,
      "end": 34.4,
      "text": "GPU. A 24 gig Mac [music] squeezes itin, but that memory is shared with MacOS. It fits. Not the tier I would buy.",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 4,
      "start": 34.4,
      "end": 42.84,
      "text": "32 gigs of unified memory.>> [music]>> Now it makes real sense on a Mac. 48 to64, real breathing room for context and",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 5,
      "start": 42.84,
      "end": 52.16,
      "text": "agents. And a 32 gig GPU keeps thentire model accelerated. So yes, aserious 27 billion parameter model runs",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 6,
      "start": 52.16,
      "end": 62.68,
      "text": "on consumer hardware. On a Mac, 32 gigs,64 to use it like an agent. Quen 3.8 isonly the beginning. Comment the model",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 7,
      "start": 62.68,
      "end": 65.0,
      "text": "you want me to test next.",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    }
  ],
  "transcript_selection": {
    "schema": "videonote.transcript_selection.v1",
    "status": "SELECTED",
    "requested_language": "en",
    "duration": 65.0,
    "reference_path": "/Users/sagawa/AI_WORK/video_notes/url/20260910_205802__What_Will_Qwen_3.8_27B_Run_On_Mac_Mini_Mac_Studio_GPUs_Tested/reference_subtitle.vtt",
    "reference_kind": "automatic_original",
    "reference_language": "en-orig",
    "selected_source": "video_source_subtitle",
    "source_caption_kind": "automatic_original",
    "source_caption_processing": "rolling_dedup_and_sentence_merge",
    "youtube_overlap_clamps": [],
    "general_high_confidence_repairs": [],
    "decision_reason": "complete_source_subtitle_fast_path_no_whisper",
    "whisper_rejected": false,
    "source_profile": {
      "segment_count": 8,
      "duration_coverage": 1.0
    }
  }
}