DeepSeek-V4.1-Flash
DeepSeek-V4.1-Flash is a natively multimodal MoE model trained from scratch on 45 trillion tokens. It uses a causal encoder-decoder layout, Compressed Sparse Attention 2, Engram conditional memory, and DSpark speculative decoding to reduce inference and KV-cache cost. Its backbone activates different parameter counts during prefill and decoding, while a separate 196B conditional-memory table is sparsely accessed. The model accepts images and text and generates text, supports million-token contexts and adjustable reasoning effort, and powers the official deepseek-flash API. MIT weights and model-specific inference code support self-hosting.
2026-09-10
552B backbone + 196B Engram memory; 8B active prefill / 16B active decode
Causal Encoder-Decoder Transformer with MoE, CSA2, Engram memory, and vision encoder
MIT
Specifications
- Parameters
- 552B backbone + 196B Engram memory; 8B active prefill / 16B active decode
- Architecture
- Causal Encoder-Decoder Transformer with MoE, CSA2, Engram memory, and vision encoder
- License
- MIT
- Context Window
- 1,000,000 tokens
- Type
- multimodal
- Modalities
- textimage
Benchmark Scores
Advanced Specifications
- Model Family
- DeepSeek
- API Access
- Available
- Chat Interface
- Not Available
Capabilities & Limitations
- Capabilities
- reasoningcode generationmathematicsvisual understandingadjustable reasoning efforttool usemillion-token context
- Known Limitations
- 552B backbone parameters exclude the additional 196B Engram memory and other components.The prefill/decode activated counts differ; sparse activation does not imply small weight-storage requirements.Deployment requires model-specific encoding and inference support.
- Notable Use Cases
- multimodal coding agentslong-context researchimage-grounded reasoning
- Function Calling Support
- Yes
- Tool Use Support
- Yes