DeepSeek officially launched the preview of its DeepSeek-V4 model series on April 24, 2026. This flagship release introduces a two-tier lineup—Pro and Flash—both of which are open-sourced under the permissive MIT license.
Key Model Specifications
The V4 series uses a Mixture-of-Experts (MoE) architecture designed for high intelligence and extreme efficiency.
- DeepSeek-V4-Pro: The flagship model with 1.6 trillion total parameters (49 billion active). It is positioned to compete with top closed-source models like OpenAI's GPT-5.5.
- DeepSeek-V4-Flash: A lighter, faster version with 284 billion total parameters (13 billion active).
- 1M Context Window: Both models support an ultra-long context length of 1 million tokens, putting them on par with major Western rivals.
Performance and Pricing
DeepSeek continues its strategy of "rock-bottom prices," claiming performance near state-of-the-art levels at a fraction of the cost.
- Benchmarking: V4-Pro leads open-source models in "world knowledge" and agentic work tasks, though it retains a high hallucination rate of approximately 94% when it lacks specific answers.
- Cost Efficiency: The Pro version is reportedly 98% cheaper than competing "Pro" models from U.S. labs. Running benchmarks on V4-Pro costs roughly $1,071, compared to over $4,800 for Anthropic's Claude 4.7 Opus.
Strategic and Hardware Shifts
A major highlight of this launch is its deep integration with domestic Chinese hardware. [13]
- Huawei Optimization: DeepSeek spent months optimizing the V4 stack for Huawei Ascend chips, marking a significant shift toward self-reliance amid ongoing U.S. chip export restrictions.
- Architectural Upgrades: Includes a new Hybrid Attention Architecture and the Muon Optimizer for faster convergence and improved stability during training.
Where to Access
- Web & App: Available via "Expert Mode" (Pro) and "Instant Mode" (Flash) at chat.deepseek.com.
- API: The updated API is live and OpenAI-compatible.
- Local Deployment: Model weights are available for download on Hugging Face and supported by tools like Ollama.
Would you like to see a benchmark comparison between DeepSeek-V4 and other recent models like GPT-5.5 or Claude 4.7?
