# DeepSeek-V4.1-Flash MoE Model Cuts Active Parameters and KV Cache for Frontier Performance

> The 552B-parameter model activates 8B parameters on input and 16B on output while supporting 1M-token context and native image understanding, with routing changes for prior versions.

*Published 2026-09-10 · By Marcus Vance*

DeepSeek-V4.1-Flash is a 552B-parameter multimodal MoE model that employs a Causal Encoder-Decoder architecture activating 8B parameters on input and 16B on output.

The introduction of DeepSeek-V4.1-Flash represents a notable step in the development of efficient frontier models by combining high performance with efficiency. DeepSeek has focused on creating an architecture that balances capability with resource efficiency. This approach addresses common challenges in large language models where high parameter counts lead to increased computational demands during both training and inference phases. The company has made the model available through its API under the identifier deepseek-flash, allowing seamless access for developers. Previous versions of the V4-Flash line have been retired in favor of this unified offering to streamline the user experience and concentrate resources on the new architecture. The move reflects a broader industry trend toward optimizing model designs for practical deployment rather than solely maximizing parameter counts. The model supports native multimodal capabilities including image understanding alongside text processing.

## What new architecture powers the DeepSeek-V4.1-Flash model?

The core innovation lies in the Causal Encoder-Decoder architecture. This design is asymmetric, meaning it uses different numbers of active parameters for the input and output phases. Specifically, it activates 8B parameters during the prefill stage for input processing and 16B parameters during the decode stage for output generation. This asymmetry allows for optimized resource use throughout the inference process. The model backbone consists of 552B parameters in total as a Mixture-of-Experts setup. Native image understanding is integrated, enabling multimodal capabilities without additional modules. The architecture is designed to support contexts of up to one million tokens. This large context window opens possibilities for handling extensive documents or long conversations in a single session. The reduction in active parameters contributes to faster inference times and higher throughput. DeepSeek has indicated that the design supports scaling to larger models in the future. The overall goal is to achieve a higher capability ceiling while improving speed and efficiency across agentic tasks.

The asymmetric nature of the encoder-decoder split targets the typical imbalance between input length and output length in many real-world applications. By dedicating fewer active parameters to prefill, the model conserves compute during the initial context ingestion phase. The decoder stage then allocates more active parameters to produce coherent responses. This structure maintains the overall 552B parameter scale while limiting runtime activation. The result is a model that delivers flagship-level agentic performance without proportional increases in memory or latency. The integration of image understanding occurs natively within the same framework, allowing unified processing of visual and textual inputs.

## How does the KV cache footprint compare to previous models?

One of the standout features is the global KV cache footprint reduced to 890 bytes per token. This represents approximately one quarter of the footprint in the previous V4-Flash model. It is also reported to be 437 times smaller than the KV cache in the V1 model. Such reductions can lead to substantial savings in memory usage during inference, particularly for long contexts. This efficiency gain is critical for deploying models at scale where memory is a limiting factor. The smaller footprint directly supports the 1M-token context capability without requiring proportional hardware upgrades. Users benefit from the ability to maintain larger conversation histories or process longer documents while keeping operational costs lower.

## What benchmark results has DeepSeek-V4.1-Flash achieved?

The model has demonstrated strong performance on several specialized benchmarks. On Humanity’s Last Exam with tools, it scored 63.9. This benchmark evaluates advanced reasoning capabilities. On Terminal-Bench 4.0, the score reached 31.2, testing terminal and command line agentic tasks. Automation-Bench yielded a score of 54.8, assessing automation and workflow capabilities. These results position the model competitively among frontier offerings despite its efficiency focus. The scores were obtained through extensive testing by the development team. The use of tools in the HLE benchmark highlights the model's agentic strengths. Performance on terminal and automation tasks suggests practical applications in software engineering and process automation. The results are detailed in the company's update documentation. These benchmarks provide concrete measures of the architecture's effectiveness in real-world scenarios.

Key benchmark scores reported for DeepSeek-V4.1-FlashBenchmarkScoreSource AttributionHumanity’s Last Exam (with tools)63.9DeepSeekTerminal-Bench 4.031.2DeepSeekAutomation-Bench54.8DeepSeek

The benchmark results underscore the model's suitability for agentic workflows that require tool use and sequential decision making. The 63.9 score on Humanity’s Last Exam with tools indicates robust reasoning under structured evaluation conditions. Terminal-Bench 4.0 at 31.2 reflects competence in command-line environments that many automation pipelines rely upon. The 54.8 on Automation-Bench further demonstrates capacity for multi-step process execution. Together these metrics illustrate consistent performance across diverse evaluation suites without reliance on excessive parameter activation during inference.

> Today, we officially release the DeepSeek-V4.1-Flash model. It is the smallest model in our new architecture family, with native multimodal visual understanding. The new architecture is designed for a higher capability ceiling, faster inference, higher throughput, and scaling to larger models.DeepSeek

## What are the availability and pricing changes for users?

The model is accessible via the DeepSeek API using the name deepseek-flash. It includes native multimodal support for images. Off-peak API pricing is set at half the peak rate, with the new pricing structure taking effect on Sept 10, 2026. This pricing adjustment aims to encourage usage during lower demand periods. Legacy V4-Flash variants have been retired and their requests now route to the new model. Starting Sept 14, 2026, V4-Pro requests will be routed to V4.1-Flash at the Flash rates. This routing will continue until the launch of V4.1-Pro. The company has stated that V4.1-Flash outperforms V4-Pro across metrics including performance, cost, speed, and total runtime based on their testing. This transition provides users with improved efficiency without additional action. The changes consolidate the product line while delivering measurable gains in operational metrics.

## What timeline governs the rollout and routing changes?

The rollout follows a structured sequence of dates and transitions. The new model became available on Sept 10, 2026. Off-peak rates apply at 50 percent of peak pricing from that date. V4-Pro requests begin routing to V4.1-Flash on Sept 14, 2026. Requests use Flash rates until the V4.1-Pro launch. Future updates may include additional architecture family members. This ordered progression ensures users experience continuous service with enhanced model capabilities. The retirement of legacy variants reduces maintenance overhead for the provider while directing traffic to the optimized release.

- The new model became available on Sept 10, 2026.
- Off-peak rates apply at 50 percent of peak pricing.
- V4-Pro requests begin routing to V4.1-Flash on Sept 14, 2026.
- Requests use Flash rates until the V4.1-Pro launch.
- Future updates may include additional architecture family members.

## What implications does this release have for the AI market and stakeholders?

The release signals a trend toward more efficient model designs in the frontier space. By reducing active parameters and KV cache requirements, DeepSeek offers a model that can deliver high performance at lower operational costs. Stakeholders in enterprise AI may find this attractive for scaling agentic applications. The multimodal capabilities expand potential use cases to include visual data processing alongside text. Market observers note that such efficiencies could influence pricing strategies across providers. For developers and businesses relying on API access, the pricing changes and performance improvements offer immediate benefits. The ability to handle 1M token contexts with reduced memory footprint supports more complex workflows. This could accelerate adoption in areas requiring long-context reasoning and automation. The retirement of older variants streamlines the product lineup for users. Enterprise teams can now allocate resources toward application development rather than infrastructure scaling.

## What developments can be expected next from DeepSeek?

DeepSeek has outlined plans for further models in the architecture family. The V4.1-Pro version is anticipated as the next major release. Until that time, the routing of V4-Pro to the Flash model ensures continuity with enhanced capabilities. The company continues to emphasize the architecture's potential for higher capability ceilings and scaling. Users can expect ongoing updates to the API and model offerings. The focus on efficiency suggests future iterations will build on the Causal Encoder-Decoder foundation. This could lead to even larger models with manageable resource demands. The emphasis on agentic performance indicates continued investment in benchmarks like those reported. Overall, the release positions DeepSeek as a key player in delivering practical frontier model advancements that prioritize both capability and deployability.

## Sources

1. [DeepSeek-V4.1-Flash is a 552B-parameter MoE model with new Causal Encoder–Decoder architecture activating 8B for input and 16B for output.](https://api-docs.deepseek.com/news/news260910)
2. [The model scores 63.9 on HLE with tools, 31.2 on Terminal-Bench 4.0, and 54.8 on Automation-Bench.](https://api-docs.deepseek.com/updates/)
3. [The model activates only 8B parameters per token during prefill and 16B during decode with global KV cache footprint to 890 bytes per token.](https://huggingface.co/deepseek-ai/DeepSeek-V4.1-Flash)

---
Source: https://aiintelreport.com/frontier-models/deepseek-v41-flash-efficient-moe-architecture
Index: https://aiintelreport.com/llms.txt · Full text: https://aiintelreport.com/llms-full.txt