DeepSeek
DeepSeek-V3
Open-weight MoE model trained at low reported cost.
Specs
- Provider
- DeepSeek
- Family
- DeepSeek
- Context window
- 128,000 tokens
- Max output
- 8,192 tokens
- Input price
- $0.27 / 1M
- Output price
- $1.1 / 1M
- Modalities
- text
- License
- deepseek
- Released
- Dec 26, 2024
- Training cutoff
- Jul 1, 2024
Estimated footprint
Per single text prompt (token-scaled rows assume ~1,000 tokens).
Footprint numbers are rough public estimates that vary by 10–100× across sources and depend on data-centre efficiency and local grid mix. They are shown as a low–high band with a source and date. Don't treat them as measured fact.
Energy
≈ 1 Wh
range 0.3–2.8 Wh
~7 s of a 9 W LED bulb
Third-party estimate (open-weight, host-dependent) · low confidence
Water
≈ 1.1 mL
range 0.35–3 mL
~0.0 s of a shower
Third-party estimate (open-weight, host-dependent) · low confidence
CO₂e
≈ 0.6 g
range 0.2–1.8 g
~5 m driven in a petrol car
Third-party estimate (open-weight, host-dependent) · low confidence
Metric sources
| Metric | Low–Mid–High | Basis | Confidence | As of | Source |
|---|---|---|---|---|---|
| co2 | 0.2 · 0.6 · 1.8 g | per prompt | low | Jan 1, 2025 | Third-party estimate (open-weight, host-dependent) |
| energy | 0.3 · 1 · 2.8 Wh | per prompt | low | Jan 1, 2025 | Third-party estimate (open-weight, host-dependent) |
| water | 0.35 · 1.1 · 3 mL | per prompt | low | Jan 1, 2025 | Third-party estimate (open-weight, host-dependent) |
Changelog
- Oct 6, 2026Seeded into PromptLedger reference library.