DeepSeek Unveils V4 Preview: A Trillion-Parameter Open-Source Leap Into the Million-Context Era

News Summary

On April 24, 2026 (Beijing Time CST), DeepSeek officially released and open-sourced a preview of its next-generation large language model series, DeepSeek-V4, marking a significant milestone in the open-source AI landscape. The release introduces two Mixture-of-Experts (MoE) variants — DeepSeek-V4-Pro and DeepSeek-V4-Flash — both featuring a standardized 1 million token context window, available immediately via DeepSeek's API, official web interface, and Hugging Face under the MIT license. The V4-Pro model, with 1.6 trillion total parameters, claims leading performance among open-source models on agentic coding, mathematics, and STEM benchmarks, while V4-Flash provides a leaner, cost-optimized alternative targeting high-frequency production workloads.

Two Models, One Architecture Leap

DeepSeek-V4 arrives as a two-tier series engineered to serve different performance and cost requirements. The flagship DeepSeek-V4-Pro carries approximately 1.6 trillion total parameters, of which only 49 billion are activated per inference token — a design characteristic of the Mixture-of-Experts architecture that keeps per-query compute tightly controlled despite the model's massive overall scale. The lighter DeepSeek-V4-Flash features 284 billion total parameters with 13 billion activated per token, prioritizing response speed and pricing efficiency for everyday and agent-oriented tasks. Both models ship with identical 1 million token context windows as a standard, default capability across all official DeepSeek services starting from the day of release (April 24, 2026 CST).

Technical Architecture: DSA, Hybrid Attention, and the Muon Optimizer

The V4 family introduces several architectural innovations that distinguish it from its predecessor, DeepSeek-V3.2. Central among these is a novel DSA (Dynamic Sparse Attention) mechanism that compresses attention computations at the token level, dramatically reducing the memory and computational costs typically associated with processing ultra-long contexts. Complementing this is a Hybrid Attention Architecture that combines local and global attention patterns — enabling the model to sustain coherent memory across sequences approaching one million tokens without the exponential cost blowup seen in traditional dense attention. The training process employs Manifold-Constrained Hyper-Connections (mHC) for stable residual signal propagation, alongside the Muon optimizer, which addresses gradient instability challenges that have historically made training trillion-parameter MoE models difficult to bring to reliable convergence. Pre-training was conducted on over 32 trillion tokens using an FP4 and FP8 mixed-precision strategy.

Benchmark Performance and Agentic Capabilities

DeepSeek released internal benchmark comparisons alongside the preview launch. On MMLU-Pro, V4-Pro was reported matching OpenAI's GPT-5.4 and trailing slightly behind Google's Gemini-3.1-Pro. In Agentic Coding evaluations, V4-Pro claimed the top position among all publicly evaluated open-source models and has reportedly become the primary agentic coding tool used by DeepSeek's own internal engineering teams. On competitive programming benchmarks such as Codeforces, V4-Pro posted scores that rival leading closed-source frontier models. Long-context retrieval benchmarks at the 1M token scale showed competitive but not leading results, with top closed-source offerings still holding advantages on some long-context memory tests. DeepSeek noted that these are preview-stage numbers and independent third-party evaluation has not yet been published.

API Availability and Developer Access

The V4 series is immediately accessible through DeepSeek's official API using the model identifiers deepseek-v4-pro and deepseek-v4-flash. The API supports both standard and reasoning (Thinking Mode) configurations, with a new reasoning_effort parameter exposing adjustable thinking intensity levels (high/max) for complex tasks such as code generation and multi-document processing. Open-source model weights are distributed via Hugging Face and ModelScope under the MIT license, allowing developers to download, run locally, and modify the models. DeepSeek announced that the legacy model identifiers deepseek-chat and deepseek-reasoner — previously mapped to V3.2 — will be deprecated on July 24, 2026 (CST), giving developers a three-month migration window.

Multimodal Roadmap and Chip Infrastructure

The current V4 preview release is text-only. DeepSeek stated in its Hugging Face announcement that the team is actively working on incorporating multimodal capabilities — meaning future versions are expected to handle image and video inputs alongside text. Separately, Huawei confirmed on April 24, 2026 (CST) that its Ascend supernode cluster, powered by Ascend 950 AI processors, fully supports the DeepSeek-V4 model series. The precise proportion of Huawei Ascend chips versus other hardware used during V4's training was not disclosed by DeepSeek in the preview release documentation.

Open-Source Strategy and Fundraising Context

DeepSeek has consistently pursued an open-source model distribution strategy, and V4 continues that approach under the permissive MIT license. The company is simultaneously reported to be pursuing its first external fundraising round, with a target valuation exceeding $20 billion. The V4 launch positions DeepSeek as a direct open-source competitor to proprietary frontier AI systems, with its cost-efficiency thesis — large total parameter counts paired with small activated parameter fractions — remaining central to its commercial and developer appeal. The preview label signals that a final, fully validated release with complete training disclosures is still forthcoming, with no strict finalization timeline announced alongside the April 24 drop.