DeepSeek Releases V4-Pro and V4-Flash Open Weights with 1M Context

DeepSeek-V4-Pro is a Mixture-of-Experts language model with 1.6 trillion total parameters and 49 billion active parameters that supports a native context of one million tokens under an MIT license.

On April 24, 2026, DeepSeek announced the preview of its V4 series including two strong Mixture-of-Experts language models. The announcement highlighted the open sourcing of the weights for both DeepSeek-V4-Pro and DeepSeek-V4-Flash. This move comes as the industry sees increasing interest in long context capabilities that allow processing of extensive documents or code repositories in a single pass. The models are positioned to offer frontier rivaling performance at lower inference costs through their architecture and licensing choices. The release includes DeepSeek-V4-Pro and DeepSeek-V4-Flash. They are both Mixture-of-Experts architectures. The release is under the MIT license. This allows broad use and modification without many restrictions. The weights are hosted on Hugging Face for easy access by developers and researchers worldwide.

The two models differ in scale but share the same context length and licensing terms. DeepSeek-V4-Pro targets users needing maximum capacity while DeepSeek-V4-Flash offers a more compact option for resource constrained environments. Both incorporate hybrid attention mechanisms that maintain output quality across the full one million token window. The MIT license means users can deploy these models in commercial products or modify the weights for specialized tasks without additional permissions.

What is the background context for the DeepSeek-V4 release?

The development of long context models has been a focus in recent years as users seek to handle more complex inputs without breaking them into smaller segments. Models with larger context windows allow for better handling of complex tasks such as summarizing entire research papers or maintaining coherence across multi hour conversations. These tasks include analyzing entire books or long code bases where previous models required multiple calls or summarization steps that could introduce errors. DeepSeek's release adds to the options available in the open source community and builds on prior work in efficient attention mechanisms.

Previously many high performance models were available only through closed APIs which limited customization and increased costs for users who needed to process sensitive data locally. The shift to open weights with permissive licenses changes this dynamic by allowing researchers and developers to fine tune and deploy locally without ongoing usage fees. The efficiency in inference is a key factor because long context often comes with high computational demands that can make deployment impractical. DeepSeek has addressed this with its architecture choices that reduce both FLOPs and memory requirements.

The AI field has seen rapid progress in model capabilities over the past several years. Context length is one area where improvements have direct user benefits because longer inputs reduce the need for chunking strategies that can disrupt context. Being able to input more information at once reduces the need for chunking strategies. This can improve the coherence of model outputs in applications ranging from legal document review to software engineering assistance.

What are the detailed specifications of DeepSeek-V4-Pro and DeepSeek-V4-Flash?

DeepSeek-V4-Pro stands out with its scale. It has a total of 1.6 trillion parameters. During operation it activates 49 billion parameters. This selective activation is typical of MoE designs and balances performance with resource use. The large total parameter count provides capacity for complex reasoning while the sparse activation keeps inference costs manageable even at the full one million token context length.

DeepSeek-V4-Flash is the lighter variant in the series. It has 284 billion total parameters. It activates 13 billion parameters during each forward pass. Despite the smaller size it maintains the same context length of one million tokens as the Pro version. This makes it suitable for scenarios where speed and lower memory footprint are prioritized over maximum parameter count.

Both models have the context length set to one million tokens by default. This is a significant increase from many existing models that top out at much shorter windows. The native support means no additional techniques are needed for this length which simplifies implementation for users who want to experiment immediately. The weights are released under the MIT license for both variants which supports broad adoption.

The license is MIT for both models. This is one of the most permissive open source licenses available. It allows anyone to use copy modify and distribute the software with minimal conditions attached to this freedom. The combination of large scale MoE design and permissive licensing distinguishes the release from many prior frontier model announcements.

What technical features enable the long context performance?

The V4 series incorporates a hybrid attention architecture. This includes Compressed Sparse Attention. It also includes Heavily Compressed Attention. These components are designed to handle long contexts efficiently by reducing the quadratic scaling issues that typically arise with standard attention mechanisms when context length increases.

In the 1M-token context setting the Pro model requires only 27 percent of the FLOPs for single token inference compared to the previous version. It also uses 10 percent of the KV cache. This reduction is significant for practical deployment because it lowers both compute and memory requirements substantially.

The hybrid approach allows the model to maintain quality while reducing the computational overhead associated with extended contexts. This is crucial for scaling context lengths without proportional increases in cost or hardware demands. The design choices reflect advancements in attention mechanisms that prioritize efficiency without sacrificing the ability to process long sequences.

Model	Total Parameters	Active Parameters	Context Length	License
DeepSeek-V4-Pro	1.6T	49B	1 million tokens	MIT
DeepSeek-V4-Flash	284B	13B	1 million tokens	MIT

What are the market and stakeholder implications of this release?

The availability of these models under open license could accelerate adoption among developers who previously relied on closed APIs. Developers can now experiment without API costs or rate limits. This is particularly relevant for long context applications in research and enterprise settings where data privacy concerns make local deployment preferable.

Stakeholders in the AI ecosystem may see changes in how models are distributed and monetized. Closed API providers might face pressure to match the context lengths or costs offered by open alternatives. Open models provide transparency that closed ones do not allowing independent verification of capabilities and potential biases.

The lower inference cost opens possibilities for more frequent use in production environments. Companies can integrate these models into their systems with less concern for compute expenses over time. This could lead to innovation in areas like document analysis and multi step agent workflows that benefit from extended context retention.

How have experts and the community reacted to the V4 models?

The release has been noted for its technical achievements in efficiency at scale. The combination of large parameter counts with MoE sparsity and long context is notable in the current landscape. The open nature encourages community contributions and fine tuning efforts on platforms that host the weights.

We present a preview version of DeepSeek-V4 series, including two strong Mixture-of-Experts (MoE) language models — DeepSeek-V4-Pro with 1.6T parameters (49B activated) and DeepSeek-V4-Flash with 284B parameters (13B activated) — both supporting a context length of one million tokens.DeepSeek-AI, Model developers

Community members on platforms like Hugging Face can access and test the models directly without barriers. This immediate availability fosters rapid feedback and iteration from a global user base. The MIT license further supports this collaborative environment by removing legal hurdles to experimentation and redistribution.

What developments are anticipated in the coming months for these models?

As a preview further refinements may be released based on community input and internal testing. The community may develop fine tuned versions for specific tasks such as code completion or multilingual processing. Integration with existing frameworks is likely to follow quickly given the standard model format on Hugging Face.

The focus on cost effective long context could influence other developers to pursue similar architectural directions. More models with comparable specifications may emerge in the open ecosystem. This would continue the trend toward open and accessible AI tools that lower barriers for independent research and application development.

Users are encouraged to explore the models on Hugging Face where the documentation provides details on usage and licensing. This release represents a step in making advanced AI more widely available to a broader range of practitioners beyond those with access to closed frontier systems.

1. Release of preview models on April 24, 2026, with detailed parameter counts and context specifications.
2. Publication of weights on Hugging Face under MIT license allowing unrestricted commercial and research use.
3. Incorporation of hybrid attention architecture including Compressed Sparse Attention and Heavily Compressed Attention.
4. Achievement of reduced FLOPs to 27 percent and KV cache to 10 percent compared with DeepSeek-V3.2 at full context length.
5. Potential for increased adoption in long-context applications across research enterprise and developer communities.