DeepSeek-V4.1-Flash
Multimodal MoE model with efficient KV cache compression
DeepSeek-V4.1-Flash is a multimodal mixture-of-experts language model released by DeepSeek AI. It features 552 billion backbone parameters and supports context windows up to one million tokens. The model uses a causal encoder-decoder architecture to reduce memory usage while processing text and images.
- 552 billion backbone parameters with sparse activation
- Supports context windows up to one million tokens
- Reduces KV cache footprint to 890 bytes per token
- Processes both text and images natively
- Activates only 8B parameters during prefill
What it is
DeepSeek-V4.1-Flash is a large language model designed for high efficiency in agentic tasks. Developed by DeepSeek AI, it combines text and image processing capabilities within a single framework. The model contains 552 billion total parameters but activates only a small fraction during operation. It supports context lengths reaching one million tokens, allowing it to handle extensive documents or long conversation histories. Its primary innovation lies in reducing the memory required for key-value caches, making it cost-effective for input-heavy workloads.
How it came about
This model follows the DeepSeek-V4-Flash lineage, aiming to improve upon its predecessor’s memory footprint. The developers trained it from scratch on a massive multimodal corpus. The training process involved extending context capabilities progressively, starting at 64,000 tokens and scaling up to one million. Post-training followed standard supervised fine-tuning and reinforcement learning protocols, with a focus on synthesising agent tasks automatically.
Inside the model
The architecture uses a Causal Encoder-Decoder design with 40 transformer layers. Half serve as a causal encoder, and half as a decoder. This structure allows the decoder to project its global key-value cache from encoder states, activating just 8 billion parameters during prefill and 16 billion during decoding. It employs Compressed Sparse Attention 2 to share attention indices across layers. FP4 quantisation further shrinks the cache size to 890 bytes per token. A vision encoder converts images into embeddings for joint processing with text.
What it can do
The model performs strongly on coding and reasoning benchmarks. It achieves a score of 79.4 on HumanEval and 93.0 on GSM8K. In agentic tests, it scores 90.6 on Terminal-Bench 2.1 and 74.2 on DeepSWE v1.1. Visual question answering results include 95.6 on DocVQA. Users can adjust reasoning effort via an integer setting from 1 to 100, balancing accuracy against inference cost. It handles tool use and complex multi-step tasks effectively.
How to use it
DeepSeek releases the model weights openly. Developers can access the model through Hugging Face. The release includes a Python reference implementation for prompt encoding, handling multi-turn conversations and tool calls. A Rust library with Python bindings is also available for production environments. This toolkit manages message formatting and API responses without requiring a Jinja chat template.
Questions people ask about DeepSeek-V4.1-Flash
How many parameters does DeepSeek-V4.1-Flash have?
The model has 552 billion backbone parameters. However, it uses a mixture-of-experts design that activates only 8 billion parameters during prefill and 16 billion during decoding, ensuring high efficiency.
Can this model process images?
Yes, it is a multimodal model. It includes a vision encoder that converts images into embeddings. These are processed jointly with text embeddings from the start of pre-training, allowing native image and text understanding.
What is the maximum context length?
The model supports context windows of up to one million tokens. This capability was developed by training on sequences of 64,000 tokens first, then extending to the full million-token length during later training stages.
How does it save memory compared to previous versions?
It uses Compressed Sparse Attention 2 and FP4 quantisation. These techniques reduce the global key-value cache footprint to roughly one-quarter of the previous DeepSeek-V4-Flash model, saving significant storage and memory.
Good reading and sources
There is no English Wikipedia article on DeepSeek-V4.1-Flash yet, so the facts on this page come from the sources listed above: huggingface.co/deepseek-ai/DeepSeek-V4.1-Flash. Some are the subject's own pages, so its claims about itself are reported as its claims. EveryLLM wrote the text in its own words, with AI assistance, and compares it with the sources to make sure no passage is copied.
It may contain mistakes, and figures change over time. Nothing here is investment or buying advice. If something is wrong, tell us. Last refreshed 2026-11-14.