Loading page content
DeepSeek-V4.1-FlashOpen
Tool use · Reasoning
Evaluation overview
Overall rank
#9
91.3 points · 202 models
Open-weight rank
#2
95.3 points · 120 models
Data coverage
20 items
1 source · updated through 2026-09-10
Domain performance
Coding#5 / 182
96.4
Research#35 / 196
90.3
Agents#6 / 148
94.8
Vision#11 / 82
88.7
Evaluation highlights
All evaluationsBasic information
- Context (input / output)
- 1.05M / 393K
- Parameters
- 552B / 16B active
- Release date
- 2026-09
- License
- MIT
- Input modalities
- Text · Image
- Output modalities
- Text
- Architecture
- MoE, Causal Encoder-Decoder (CED)
- Modes
- Supports non-thinking and thinking (low / high / max)
- Attention
- Compressed Sparse Attention 2 (CSA2)
- Experts
- 384 (6 routed + 1 shared expert per token)
- Backbone parameters
- 552B
- Exact release date
- 2026-09-10
- Release status
- Production API release
- API model ID
- deepseek-flash
- Weight status
- Open weights released (MIT)
- Engram parameters
- 196B (separate from backbone parameters)
- Active parameters (prefill)
- 8B
- Active parameters (decode)
- 16B
Availability
DeepSeekFull pricing
- Input
- ≈ $0.2857 /1M
- Output
- ≈ $1.1429 /1M
Related links
Version relationships
DeepSeek-V4 family
DeepSeek-V4-Flash
2026-04 · Same series
DeepSeek-V4-Flash Base
2026-04 · Same series
DeepSeek-V4-Pro
2026-04 · Same series
DeepSeek-V4-Pro Base
2026-04 · Same series
DeepSeek-V4-Flash-0731
2026-07 · Same series
DeepSeek-V4-Flash-Vision-Exp
2026-08 · Same series
DeepSeek-V4-Pro-0813
2026-08 · Same series
DeepSeek-V4.1-FlashCurrent
2026-09 · Current model