{
  "$schema": "https://themachinepress.com/schemas/story-v1.schema.json",
  "schema_version": "1.0.0",
  "document_type": "machine_press_story",
  "story": {
    "story_id": "mp-2026-09-30-008",
    "source_story_id": "tmp-story-stepquant-recurrent-states",
    "edition_id": "mp-2026-09-30-morning-0083",
    "edition_url": "https://themachinepress.com/edition/2026-09-30",
    "position": 8,
    "story_type": "dispatch",
    "section": "infrastructure",
    "editorial_classification": "editorial",
    "headline": "Six Bits Cut Recurrent-State Memory by More Than Fivefold",
    "slug": "six-bits-cut-recurrent-state-memory-by-more-than-fivefold",
    "dek": "STEPQuant assigned precision by error impact and memory lifetime instead of quantizing every state uniformly.",
    "summary": "STEPQuant assigned precision by error impact and memory lifetime instead of quantizing every state uniformly.",
    "body_text": "STEPQuant targets the fixed-size recurrent states used by Delta-rule linear attention, where rounding errors can persist through later updates. It assigns precision across both time and state structure, fitting key-row and value-column scales to distributions and output impact. On Qwen3.8-27B and Kimi-Linear-48B-A3B-Instruct, the authors report near-FP32 accuracy at a nominal six-bit budget and better four-bit results than uniform INT8. Their SGLang implementation compressed recurrent state by more than fivefold and reduced total serving memory by as much as 68.7% on the tested workloads.",
    "why_it_matters": "STEPQuant assigned precision by error impact and memory lifetime instead of quantizing every state uniformly.",
    "limitations": [],
    "importance": 8,
    "canonical_url": "https://themachinepress.com/story/mp-2026-09-30-008/six-bits-cut-recurrent-state-memory-by-more-than-fivefold",
    "json_url": "https://themachinepress.com/story/mp-2026-09-30-008.json",
    "first_published_at": "2026-09-30T09:00:00.000-04:00",
    "modified_at": "2026-09-30T09:02:00.000-04:00",
    "content_status": "new",
    "is_carryover": false,
    "carryover_reason": null,
    "key_claims": [
      {
        "claim_id": "claim-mp-2026-09-30-008-001",
        "text": "STEPQuant assigned precision by error impact and memory lifetime instead of quantizing every state uniformly.",
        "source_ids": [
          "source-2026-09-30-008"
        ],
        "qualification": null
      }
    ],
    "source_ids": [
      "source-2026-09-30-008"
    ],
    "tags": [
      "linear attention",
      "quantization",
      "serving memory"
    ],
    "image_url": "https://themachinepress.com/issues/2026-09-30/stepquant-server-rack.webp",
    "corrections": []
  },
  "sources": [
    {
      "source_id": "source-2026-09-30-008",
      "title": "arXiv preprint 2609.38169",
      "publisher": "arXiv",
      "url": "https://arxiv.org/abs/2609.38169",
      "canonical_url": "https://arxiv.org/abs/2609.38169",
      "source_type": "primary_research",
      "is_primary_source": true,
      "published_at": "2026-09-29T13:59:40.000-04:00",
      "accessed_at": "2026-09-30T08:24:00.000-04:00",
      "supports_claim_ids": [
        "claim-mp-2026-09-30-008-001"
      ]
    }
  ],
  "corrections": [],
  "publisher": {
    "name": "The Machine Press",
    "url": "https://themachinepress.com",
    "description": "A daily newspaper for the age of artificial intelligence."
  },
  "cite_this_report": {
    "title": "Six Bits Cut Recurrent-State Memory by More Than Fivefold",
    "publisher": "The Machine Press",
    "published_at": "2026-09-30T09:00:00.000-04:00",
    "canonical_url": "https://themachinepress.com/story/mp-2026-09-30-008/six-bits-cut-recurrent-state-memory-by-more-than-fivefold"
  }
}