deepseek-ai/DeepSeek-V4.1-Flash
DeepSeek-V4.1-Flash is a 763 billion parameter multimodal Mixture-of-Experts (MoE) model developed by DeepSeek-AI, supporting contexts up to one million tokens. It features a Causal Encoder-Decoder (CED) architecture and Compressed Sparse Attention 2 (CSA2) for significant KV cache compression, reducing the global KV cache footprint to 890 bytes per token. This model is optimized for cost-efficient agentic workloads, natively processing images and text, and excels in agentic benchmarks and code-related tasks.
Loading preview...
DeepSeek-V4.1-Flash: KV Cache Compression and Multimodal Agentic Performance
DeepSeek-V4.1-Flash is a powerful multimodal Mixture-of-Experts (MoE) model from DeepSeek-AI, featuring 552 billion backbone parameters and supporting an extensive context window of up to one million tokens. It is designed for efficiency and advanced agentic capabilities, natively handling both image and text inputs.
Key Innovations & Capabilities
- Causal Encoder-Decoder (CED) Architecture: This unique architecture projects the decoder's global KV cache from the encoder's final hidden states, allowing for highly efficient parameter activation (8B during prefill, 16B during decode) crucial for input-heavy agentic tasks.
- Compressed Sparse Attention 2 (CSA2): Employs a novel attention mechanism that assigns static modes (Full, Reindex, Reuse) to layers, combined with a Hierarchical Sparse Indexer and FP4 main KV caching. These innovations drastically reduce the global KV cache footprint to just 890 bytes per token, a 4-fold reduction compared to DeepSeek-V4-Flash.
- Multimodal Processing: Integrates a custom DeepSeek-ViT vision encoder, enabling joint processing of visual and text embeddings from the start of pre-training.
- Agentic Performance: Demonstrates strong performance across various agentic benchmarks, including significant improvements in Terminal-Bench 2.1, DeepSWE v1.1, and AutomationBench, often outperforming other frontier models.
- Controllable Reasoning Effort: Supports a continuously controllable reasoning effort setting (1-100), allowing users to trade inference cost for accuracy.
- Extensive Pre-training: Trained from scratch on a massive multimodal corpus of 45 trillion tokens, with sparse attention trained at 64K sequence length and context extended to 1M tokens.
Ideal Use Cases
DeepSeek-V4.1-Flash is particularly well-suited for:
- Agentic Workloads: Its optimized architecture and strong benchmark results make it excellent for complex agent-based applications requiring efficient processing of long inputs.
- Multimodal Applications: For tasks that require understanding and generating content from both images and text.
- Cost-Sensitive Deployments: The significant KV cache compression offers substantial cost savings for high-volume inference.
- Code Generation and Reasoning: Achieves competitive scores on code and math benchmarks like HumanEval and GSM8K.