DeepSeek has launched DeepSeek V4.1 Flash, a new multimodal AI model designed to improve inference efficiency, throughput and cost while supporting large-scale agentic workloads. Announced on September 10, 2026, the model is the smallest member of DeepSeek’s new architecture family and combines a 552-billion-parameter Mixture-of-Experts (MoE) backbone with an asymmetric Causal Encoder-Decoder design.
Unlike conventional models where the same architecture processes input and output, V4.1 Flash activates 8 billion parameters during input processing and 16 billion during decoding. DeepSeek says the design is intended to deliver higher capability without requiring the full model to be active for every token.
The release also brings native visual understanding, a context window of up to 1 million tokens and a substantially smaller KV cache. These changes are particularly relevant to AI agents, where long contexts and repeated inputs can make inference expensive.
DeepSeek V4.1 Flash Uses an Asymmetric 552B MoE Architecture
At the center of V4.1 Flash is a 552B-parameter Mixture-of-Experts model.
Rather than activating the entire parameter set for every token, the model uses a Causal Encoder-Decoder architecture designed to separate input processing from output generation.
DeepSeek’s technical report describes a 40-layer Transformer divided into a 20-layer causal encoder and a 20-layer decoder. This allows the model to activate approximately 8B parameters per token during prefill and 16B during decoding.
This architecture is particularly suited to workloads where applications repeatedly send large amounts of input context. That includes coding agents, research agents, enterprise automation and other systems that process long-running conversations or documents.
Key DeepSeek V4.1 Flash specifications
| Specification | DeepSeek V4.1 Flash |
|---|---|
| Total backbone parameters | 552B |
| Active parameters | 8B input / 16B output |
| Architecture | Causal Encoder-Decoder |
| Model type | Mixture of Experts |
| Context window | Up to 1M tokens |
| Vision | Native multimodal |
| KV cache | About 1/4 of V4-Flash |
| License | MIT |
| API model name | deepseek-flash |
DeepSeek’s model card confirms that the model supports text and images natively and is released under an MIT license.
Smaller KV Cache Is a Major Part of the DeepSeek V4.1 Flash Design
One of the most important technical changes in DeepSeek V4.1 Flash is its approach to KV caching.
KV, or key-value, caching stores information from previous tokens so the model does not have to repeatedly recompute the same context. For long-context applications and AI agents, this can become a major infrastructure cost.
DeepSeek says V4.1 Flash requires approximately one-quarter of the HBM and one-eighth of the SSD storage used by the previous generation for its KV cache.
The technical report attributes the reduction to several architectural techniques, including Causal Encoder-Decoder processing, SWA Bounded Replay, Compressed Sparse Attention 2 and FP4 KV caching. The reported global KV cache footprint is approximately 890 bytes per token.
This matters because AI agents can repeatedly process large contexts. Reducing persistent cache requirements can lower memory pressure and potentially reduce the infrastructure required to serve those workloads.
Native Multimodal AI Comes to V4.1 Flash
DeepSeek V4.1 Flash also introduces native visual understanding as part of the model.
The Hugging Face model card describes the system as a multimodal MoE model capable of processing both images and text. It supports contexts of up to 1 million tokens.
That expands the model beyond conventional text-only large language model workloads.
Potential applications include:
- Image and document analysis
- Coding and software development
- Long-context research
- AI agents
- Enterprise document processing
- Multimodal automation
- Visual reasoning workflows
The combination of multimodal input and long context could make the model useful for agents that need to work across documents, screenshots, code and other forms of information.
DeepSeek V4.1 Flash Targets AI Agent Workloads
The architecture is particularly relevant to the growth of AI agents.
Agentic systems often generate many intermediate inputs, repeatedly access context and use external tools. As those workflows become longer, the cost of processing and storing context becomes increasingly important.
DeepSeek explicitly links its KV-cache improvements to agent workloads, arguing that cache-related costs can represent a significant portion of agent usage.
The model also supports tool calls, JSON output and both OpenAI-compatible and Anthropic-compatible API interfaces through DeepSeek’s API. The current API documentation lists a 1-million-token context window and a maximum output of up to 384K tokens.
DeepSeek V4.1 Flash Brings Lower API Pricing
DeepSeek has also positioned V4.1 Flash as a lower-cost model for high-volume inference.
The current API pricing lists V4.1 Flash at:
| Pricing | Off-Peak | Peak |
|---|---|---|
| 1M input tokens, cache hit | $0.003 | $0.006 |
| 1M input tokens, cache miss | $0.15 | $0.30 |
| 1M output tokens | $0.60 | $1.20 |
DeepSeek says off-peak rates are half the peak rates. Peak hours are currently listed as 01:00–04:00 and 06:00–10:00 UTC on weekdays.
The pricing structure could be particularly relevant for developers operating large AI automation workloads, where small reductions in per-token costs can become significant at scale.
DeepSeek Is Retiring Earlier Flash Models
V4.1 Flash has also changed the company’s model lineup.
DeepSeek says the previous V4 Flash and V4 Flash Vision Experimental models have been retired. Legacy model names remain accepted for compatibility, but requests are served by V4.1 Flash and billed at the Flash rate.
DeepSeek initially announced plans to phase out V4 Pro as well, but its current API documentation now says that V4 Pro API service will continue after September 14, 2026, with its existing billing method.
That distinction is important for developers maintaining production applications because model-routing announcements can directly affect API compatibility and operating costs.
Open-Weight Deployment Expands DeepSeek’s Reach
DeepSeek has also released V4.1 Flash through Hugging Face under an MIT license.
The model can be used with frameworks including Transformers, vLLM and SGLang, while the model page also points to quantization options and local deployment tools.
DeepSeek says it plans to work with the open-source community on inference support and explore additional deployment options. The company specifically references large-scale deployments involving thousands of GPUs and storage infrastructure.
This gives developers and organizations more options beyond using the hosted API.
Why DeepSeek V4.1 Flash Matters?
The significance of V4.1 Flash is less about the headline 552B parameter count and more about how much of that model needs to be activated and stored during inference.
Its design reflects a broader trend in AI development: improving model efficiency rather than simply increasing parameter counts.
For developers, the combination of:
- Large context
- Native multimodal processing
- Sparse expert activation
- Smaller KV cache
- Tool calling
- Open-weight availability
- Lower API pricing
Could make the model particularly attractive for high-volume AI agents and automation systems.
Conclusion
DeepSeek V4.1 Flash marks a significant shift toward more efficient AI model architectures. Its 552B MoE backbone, asymmetric processing, native multimodal capabilities and substantially reduced KV-cache requirements are designed to make large-scale AI inference more economical.
The model is particularly relevant to AI agents and automation systems that repeatedly process large amounts of context. Combined with open-weight availability and lower API pricing, V4.1 Flash gives developers another option for building high-volume AI applications while highlighting a broader industry trend toward improving AI performance per unit of compute, rather than relying solely on larger models.
FAQs
1. What is DeepSeek V4.1 Flash?
DeepSeek V4.1 Flash is a 552B-parameter multimodal Mixture-of-Experts AI model with an asymmetric Causal Encoder-Decoder architecture and support for contexts up to 1 million tokens.
2. How many parameters does DeepSeek V4.1 Flash activate?
The model activates approximately 8B parameters during input processing and 16B during decoding, despite having a 552B-parameter backbone.
3. Does DeepSeek V4.1 Flash support images?
Yes. V4.1 Flash natively supports multimodal input, including image and text processing.
4. What is the context window of DeepSeek V4.1 Flash?
The model supports a context window of up to 1 million tokens.
5. Is DeepSeek V4.1 Flash open source?
DeepSeek has released the model weights on Hugging Face under an MIT license and provides deployment paths through tools such as Transformers, vLLM and SGLang.
6. How much does DeepSeek V4.1 Flash API access cost?
Current DeepSeek API pricing lists V4.1 Flash at $0.15 per 1 million input tokens for cache misses and $0.60 per 1 million output tokens during off-peak periods. Peak prices are twice those rates.
Also Read –
DeepSeek Model 1: FlashMLA and Optimized Attention Explained
DeepSeek mHC: A Fundamental Shift in Transformer Architecture
