← Back to directory
DeepSeek
DeepSeek-V3
Released Dec 26, 2024 · knowledge cutoff 2024-07
- Status
- Superseded
- Location
- China
- Modality
- Text
- Context window
- 128K
- Max output
- 8K
- Speed
- —
- Price ($/MTok in / out)
- $0.27 / $1.10
- Cost per task
- $0.023
- Open weights
- Yes
Benchmarks5 of 8 reported
MMLU-Pro
75.9%
GPQA Diamond
59.1%
SWE-bench Verified
42%
AIME (latest)
39.2%
LMArena Elo
1318
Terminal-Bench 2.1
—
Humanity's Last Exam
—
ARC-AGI-2
—
Strengths & weaknesses
Strengths
- Sparse routing makes a very large model cheap to run per token
- Fully downloadable, so it can go where no external API is permitted
- Strong at code and maths relative to what it costs to serve
Weaknesses
- Text-only — no images, charts or documents
- English prose reads stilted and repetitive next to US flagships
- Deflects or refuses on politically sensitive subjects
For developers
API model strings
Not researched
Licence
Not researched
Retirement
No retirement announced
Lineage
No recorded predecessor