DeepSeek
DeepSeek V4 Flash (Non-reasoning)
DeepSeek-V4 series incorporate several key upgrades in architecture and optimization: (1) a hybrid attention architecture that combines Compressed Sparse Attention (CSA) and Heavily Compressed Attention (HCA) to improve long-context efficiency; (2) ManifoldConstrained Hyper-Connections (mHC) that enhance conventional residual connections; (3) and the Muon optimizer for faster convergence and greater training stability
DeepSeek release notesModel Snapshot
TypeLanguage
ReleaseApr 23, 2026
Context1,048,576
Tagsreasoning, tool-use, implicit-caching
Overall
Rank #172 across 600
28.7
Coding
No rank
n/a
Blended Price
Rank #47 across 379
$0.175/M
Strength Profile
Percentile score by analysis domain.
* Cost is inverted: lower input, output, and blended prices rank higher.
Benchmark Percentiles
Higher bars mean stronger relative placement.
Model Details
Runtime
Output Speed101.9 tok/s
First Token0.94s
Max Output393,216
Pricing
Blended$0.175/M
Input$0.140/M
All Benchmarks
| Metric | Domain | Value | Rank |
|---|---|---|---|
| Artificial Analysis Intelligence Index | overall | 28.7 | #172 |
| GPQA | reasoning | 71.6% | #252 |
| Humanity's Last Exam | reasoning | 7.0% | #275 |
| SciCode | coding, reasoning | 37.3% | #210 |
| Output Speed |