DeepSeek officially released preview versions of its flagship DeepSeek-V4 series on Friday, April 24, 2026. This release marks the company’s first major ground-up model launch since the industry-disrupting R1 model of 2025.
The V4 series introduces two primary variants designed to compete with the top-tier “frontier” models from OpenAI, Google, and Anthropic, but at a significantly lower price point.
1. The Lineup: Pro and Flash
The V4 series follows a Mixture-of-Experts (MoE) architecture, allowing for massive total parameter counts while maintaining high inference efficiency.
| Feature | DeepSeek-V4-Pro | DeepSeek-V4-Flash |
| Total Parameters | 1.6 Trillion | 284 Billion |
| Active Parameters | 49 Billion | 13 Billion |
| Context Window | 1 Million Tokens | 1 Million Tokens |
| Primary Goal | Frontier Reasoning & Logic | Speed and Economy |
| License | MIT License (Open Source) | MIT License (Open Source) |
- World’s Largest Open Model: With 1.6 trillion parameters, V4-Pro is now the largest open-weights model in existence, surpassing Kimi K2.6 and GLM-5.1.
- Million-Length Context: Both models utilize a new Hybrid Attention Architecture and Engram Conditional Memory to handle 1-million-token contexts without the massive computational costs traditionally associated with long sequences.
2. Benchmark Performance
DeepSeek claims that V4-Pro significantly bridges the gap between open-source and closed-source models, particularly in coding and reasoning tasks.
- Coding: On the MMLU-Pro benchmark, V4-Pro reportedly matches GPT-5.4 and sits just behind Gemini 3.1 Pro and Claude Opus 4.6.
- SWE-bench Verified: Internal claims place the model between 80-85%, which would put it in direct competition with the most advanced “agentic” coding models.
- Agentic Tasks: Optimized for use in AI Agent tools like Claude Code and OpenClaw, the model is designed to perform autonomous research and software development.
3. Disruption through Pricing
True to DeepSeek’s reputation, the V4 series is priced aggressively—up to 50x cheaper than its Western counterparts.
- V4-Flash: $0.14 per 1M tokens (Input) / $0.28 per 1M tokens (Output).
- V4-Pro: $1.74 per 1M tokens (Input) / $3.48 per 1M tokens (Output).
- Comparison: For context, V4-Pro is roughly 10x cheaper than GPT-5.4 for output tokens and nearly 20x cheaper than Claude Opus 4.6.
4. Hardware Optimization (The “China” Advantage)
Reflecting current geopolitical realities, the V4 series is the first frontier model optimized specifically for Chinese silicon alongside Nvidia hardware.
- Dual-Chip Compatibility: The software stack is optimized for both Nvidia H100/H200 and Huawei Ascend 950PR systems.
- Efficiency Gains: By utilizing Manifold-Constrained Hyper-Connections (mHC), DeepSeek has improved training stability and energy efficiency, allowing them to train at a fraction of the cost of their rivals.
