@meta
  v: 1
  route: /model/deepseek-v3
  generated: 2026-06-10T09:14:24.300Z

@intent
  purpose:    Show what hardware can run DeepSeek V3.
  audience:   ai-engineer, self-hoster, model-evaluator
  capability: inspect_model_specs, find_compatible_gpus, compare_similar_models, open_huggingface

@state
  slug: deepseek-v3
  name: DeepSeek V3
  hf_repo: deepseek-ai/DeepSeek-V3
  params_b: 671
  active_params_b: 37
  family: DeepSeek
  type: moe
  context_k: 128
  fp16_gb: 1342
  summary: "671B MoE, 37B active. GPT-4 class on a fraction of the inference cost. Multi-node serving at FP16."
  smallest_fitting_gpu_q4: m3-ultra-512
  smallest_fitting_gpu_fp16: ~

@actions
  - id: open_huggingface
    method: GET
    href: https://huggingface.co/deepseek-ai/DeepSeek-V3
  - id: view_compatible_gpus
    method: GET
    href: /model/deepseek-v3
  - id: compare_similar
    method: GET
    href: /model/deepseek-r1
  - id: view_calculator
    method: GET
    href: /#calculator

@context
  > 671B MoE, 37B active. GPT-4 class on a fraction of the inference cost. Multi-node serving at FP16.

@nav
  self:      /model/deepseek-v3
  parents:   [/]
  peers:     [/model/deepseek-r1, /model/llama-3-1-405b, /model/mixtral-8x22b, /model/gpt-oss-120b]
  drilldown: /the-math