id: moonshotai__Kimi-K3__strategic
status: approved
name: Kimi-K3
org: moonshotai
source:
  kind: huggingface
  url: https://huggingface.co/moonshotai/Kimi-K3
  revision: main
  model_card_url: https://huggingface.co/moonshotai/Kimi-K3
  project_url: https://www.kimi.com/
  paper_url: https://huggingface.co/papers/2607.24653
license:
  spdx: LicenseRef-Kimi-K3
  url: https://huggingface.co/moonshotai/Kimi-K3
  allows_offline_retention: true
  allows_local_ops: true
  allows_fine_tune: true
  notes: >
    Custom Kimi K3 License (not plain MIT/Apache). Review commercial SaaS
    revenue thresholds and attribution before external product embedding.
    Confirm card text at pin time.
profile:
  summary: "S1 open frontier MoE (2.8T) — best open general/agentic class mid-2026."
  original_source: https://huggingface.co/moonshotai/Kimi-K3
  use_cases:
    - "Long-horizon coding and agent workflows (offline optionality)"
    - "Strategic reserve if closed APIs are unavailable or restricted"
    - "Future multi-GPU / training-facility inference when hardware catches up"
    - "Capability baseline against closed frontier for daily briefs"
    - "Native vision + 1M-context knowledge work (when runnable)"
  sweet_spots:
    - "Top open-weight intelligence class after Jul 2026 weight release"
    - "Agentic / coding / long-context open MoE identity"
    - "Displaces prior S1 DeepSeek-V3 as primary open general target"
  not_ideal_for:
    - "Current single-GPU or small multi-GPU lab serve (far beyond run envelope)"
    - "Fitting under 850 GiB soft quota at full precision (~1.45 TiB tree)"
    - "Casual always-on chat (use R spine)"
  capability_notes: >
    Moonshot Kimi K3: ~2.8T total MoE parameters, sparse activation (~16/896
    experts), Kimi Delta Attention + Attention Residuals, native multimodal,
    ~1M context. Official HF tree measured ~1454 GiB (96 shards). Collection
    deferred until capacity exception or official compressed distribution.
swot:
  strengths:
    - "Leading open-weight general/agentic capability at release"
    - "Open weights enable true offline retention optionality"
    - "Strong coding / long-horizon / vision narrative"
  weaknesses:
    - "Disk footprint (~1.45 TiB) exceeds current soft quota and most lab GPUs"
    - "Custom license with commercial thresholds"
    - "Serve cost multi-node enterprise class"
  opportunities:
    - "Hold identity in catalog; pull compressed later if published"
    - "Anchor daily briefs / evals against open SOTA"
    - "Displace V3 as primary S1 when space allows"
  threats:
    - "Quota: full pull blocks all other S niches on VAULT"
    - "Rapid supersession by K3.x or peer open giants"
    - "Third-party quant quality variance if forced to unofficial packs"
size:
  total_bytes: 1560998984390
  total_human: "~1454 GiB official HF tree (measured 2026-08-03); compressed TBD"
hardware_class:
  min_vram_gb_q4: 0
  min_vram_gb_fp16: 0
  notes: "Beyond current lab run envelope (multi-node). Strategic reserve only."
axes: [A, B, C, D]
priority: high
collection:
  approved_by: "bernd"
  approved_at: "2026-08-03"
  downloaded_at: null
  downloaded_by: ""
  storage_path: ""
  brief_refs:
    - briefs/2026/08/2026-08-03.md
    - research/2026-07-24-nas-strategic-collection-plan.md
    - docs/decisions/2026-07-24-nas-strategic-reserve.md
reason: >
  S1 strategic — best open-weight general/agentic model class after Jul 2026
  release; keep capability identity even though full tree exceeds soft quota.
tags: [tier-s, strategic, beyond-run-envelope, moe, general, vision, kimi, s1]
companions: []
notes: >
  Do not bulk-download full tree under 850 GiB soft quota. Prefer official
  compressed if/when available. Revisit capacity exception with operator.
history:
  - at: "2026-08-03"
    event: nominated
    by: grok
    detail: "Week catch-up brief — Kimi K3 open weights Jul 27 as new open SOTA."
  - at: "2026-08-03"
    event: approved
    by: bernd
    detail: "Operator direction: catalog and plan; collection deferred on capacity."
  - at: "2026-08-03"
    event: collection_deferred
    by: grok
    detail: "Official tree ~1454 GiB > soft quota 850 GiB; await capacity decision or compressed release."
  - at: "2026-08-03"
    event: profile_swot_added
    by: grok
    detail: "schema 0.2 profile + SWOT."