DeepSeek · open · 2026-04
DeepSeek-V4-Flash
#10 index 85.2
the takeThe same trick as V4-Pro in a size you could almost self-host, for pennies. Same caveat too: the scores are DeepSeek's own until someone neutral runs them.
SWE-bench Verified
79.0
GPQA Diamond
88.1
MMLU-Pro
86.4
- Context
- 1M
- Input
- $0.14/M
- Output
- $0.28/M
- Speed
- —
- Modality
- text
- Among the cheapest frontier-tier APIs
- 1M context in a compact MoE
- Open weights, strong reported coding
- Vendor-reported, unverified scores
- Trails V4-Pro on the hardest tasks
- Text-only