DeepSeek-V4.1-Flash is coming to Microsoft Foundry

Agentic applications increasingly need to process entire code repositories, large document collections, images, tool results, and long-running conversation histories. As that context grows, the key-value, or KV, cache maintained during inference can become a major constraint on cost, memory, and deployment scale.

DeepSeek-V4.1-Flash is coming to Foundry Models. This multimodal Mixture-of-Experts model supports text and image inputs, a context window of up to one million tokens, and continuously configurable reasoning effort. Developers can deploy the model either Direct from Azure (public preview) or through Fireworks on Foundry (generally available), choosing the consumption path that best fits their performance, operational, and purchasing requirements.

More capability and control for agentic applications

DeepSeek-V4.1-Flash is more than a fine-tuned version of V4-Flash. It introduces a Causal Encoder-Decoder architecture in which a 20-layer causal encoder processes the input and a 20-layer decoder generates the output. The model has a 552-billion-parameter backbone, activating 8 billion parameters during prefill and 16 billion during decoding.

The model also expands beyond the text-only capabilities of V4-Flash. It adds native image understanding and replaces three fixed reasoning modes with a continuous reasoning-effort setting from 1 to 100.

Developers can use lower effort for tasks such as routing, classification, and extraction, then increase effort for planning, coding, debugging, and multi-step tool use, without changing model deployments.

Efficient long-context AI starts with a smaller memory footprint

Long-context agents need to retain code, documents, images, tool results, and conversation history during inference. As that context grows, the KV cache can become a significant memory and scaling constraint.

DeepSeek reports that V4.1-Flash requires 890 bytes of global KV cache per token, approximately one quarter of V4-Flash’s footprint. Its architecture is particularly relevant to workloads that read significantly more information than they generate.

For developers, the practical opportunity is to retain more useful context and support context-heavy workloads more efficiently. Actual latency, concurrency, infrastructure requirements, and cost will depend on the application, serving configuration, traffic profile, and deployment path.

Stronger performance on agentic workloads

DeepSeek-V4.1-Flash pairs its more efficient long-context architecture with improved performance on agentic tasks. On DeepSeek’s published evaluations, V4.1-Flash improves on V4-Flash across the reported agentic benchmarks, including coding, terminal, security, tool-use, and automation scenarios.

On DeepSWE v1.1, V4.1-Flash resolves 74.2% of tasks, compared with 54.4% for V4-Flash. DeepSeek also reports improvements on Terminal-Bench 2.1, CyberGym, AutomationBench, and Agent’s Last Exam. Results can vary based on prompts, orchestration framework, context construction, deployment, and workload. Teams should evaluate model quality, latency, and cost using their own application data. Competitor or comparative figures reported by DeepSeek have not been independently verified by Microsoft.

What this means for developers

Put together, the architecture and the benchmark profile point to a specific set of workloads where DeepSeek-V4.1-Flash is a strong fit:

• Repository-scale coding agents that load large codebases, run tools in a terminal, and iterate over many steps.


• Document-heavy and multimodal analysis, including PDFs, charts, screenshots, and scanned forms, where image inputs and a 1M-token window remove the need for aggressive chunking.


• Long-running automation and research agents that accumulate tool outputs and conversation history over hundreds of turns.


• Cost-sensitive, high-concurrency deployments where the smaller KV cache translates directly into more sessions per GPU.

The reasoning effort control deserves special attention. Rather than choosing between a fast model and a thinking model, you can run one deployment and set effort per request: low for routing and extraction, high for planning and debugging. That simplifies orchestration compared with the V4 Flash and V4 Pro pairing, and it gives you a single, continuous knob to tune against quality and latency targets in Foundry evaluations.

One model, two deployment paths

As the open model ecosystem evolves, teams increasingly want flexibility not only in which model they use, but also in how they deploy and consume it. With DeepSeek-V4.1-Flash, developers can select between two paths in Microsoft Foundry based on the needs of their application.

The right path depends on the workload. Developers should evaluate model quality, latency, throughput, deployment shape, data requirements, and cost using their own application traffic and evaluation datasets.

Other models now available through Fireworks on Foundry

Microsoft Foundry is also expanding model choice through Fireworks with additional global deployment options for leading open models. These updates give developers more flexibility to evaluate and deploy models for agentic, coding, reasoning, and long-context workloads through the same Foundry experience.

• DeepSeek-V4.1-Flash is available through Fireworks AI with PayGo Global and Provisioned Throughput deployment options.


• GLM-5.3-Flash is available through Fireworks with a Global deployment option.


• Kimi K3 now adds a Global deployment option through Fireworks, expanding beyond its previously available Data Zone deployment.

Availability, deployment options, and pricing may vary by model. Review each model card in the Microsoft Foundry model catalog for the latest details before deployment.

DeepSeek V4.1 Flash Pricing

A unified Foundry experience

Whichever deployment path teams choose, Microsoft Foundry provides a common environment for discovering models, comparing options, evaluating them against application-specific data, and managing AI development with integrated governance and observability.

This lets developers focus on selecting the right model and deployment path for their scenario instead of assembling fragmented tools and separate operational environments. Model discovery, evaluation, and management stay centered in Foundry, while the consumption path adapts to the team.

Getting started

Explore DeepSeek-V4.1-Flash in the Microsoft Foundry model catalog through either deployment path:

• DeepSeek-V4.1-Flash, Direct from Azure: DeepSeek-V4.1-Flash | Model Catalog | Microsoft Foundry


• DeepSeek-V4.1-Flash, Fireworks on Foundry: FW-DeepSeek-V4.1-Flash | Model Catalog | Microsoft Foundry

You can then:

• Compare the available deployment paths


• Evaluate the model using your own datasets and agent scaffolds, sweeping the reasoning effort setting


• Select the deployment option aligned with your operational and performance requirements


• Begin integrating the model into long-context, multimodal, and agentic applications

For the full architecture description and complete benchmark tables, see the DeepSeek-V4.1-Flash model card on Hugging Face