DeepSeek V4 Ready for Launch: March 4 Release Likely
DeepSeek is expected to unveil its next-generation V4 model on March 4, timed with the opening of China's Two Sessions. With groundbreaking specs including a 1-trillion-parameter MoE architecture, 1-million-token context window, and Engram memory technology, combined with pricing at roughly one-sixth the cost of GPT-5.2, V4 is poised to shake up the AI landscape.
March 4 Two Sessions Opening: V4 Launch Countdown
Chinese AI startup DeepSeek is expected to unveil its next-generation large language model, V4, on March 4. The date coincides precisely with the opening of the National People's Congress (Two Sessions), China's most significant political event. DeepSeek had repeatedly hinted at a V3 successor earlier this year but postponed multiple times, and now definitive launch signals are finally emerging.
According to multiple industry insiders and international media outlets, DeepSeek plans to announce V4 within this week. The timing during the Two Sessions is widely interpreted as a strategic move to showcase China's AI self-reliance on the global stage. TechNode, citing DeepSeek internal sources, reported that "a release this week is certain," while PYMNTS also conveyed an "imminent announcement" tone.
DeepSeek made waves in the open-source AI market with the December 2024 launch of V3, and sent shockwaves through the global AI industry with its reasoning-specialized R1 model in January 2025. V4 represents the next leap: a native multimodal model spanning text, images, and video.
Groundbreaking Specs: 1T MoE and 1-Million-Token Context
V4's most striking change is its massive architectural expansion. The model adopts a Mixture-of-Experts (MoE) structure with approximately 1 trillion total parameters, while keeping only about 32 billion parameters active during inference, balancing efficiency and performance. Compared to V3's 671 billion parameters (37B active), V4 dramatically scales total size while actually reducing active parameters, maximizing computational efficiency.
The context window expands to 1 million tokens, dwarfing GPT-4o's 128K tokens and Claude 3.5 Sonnet's 200K tokens. The key technology enabling this is Engram Conditional Memory. Unlike traditional transformers that exhibit O(n) or higher computational complexity when processing long contexts, Engram achieves O(1) lookup time for past context, inspired by the human brain's long-term memory mechanisms.
According to Tom's Hardware, Engram is embedded in both training and inference stages, forming "weight-based permanent memory." This means the model can efficiently leverage context from past sessions, not just long text within a single prompt. Additionally, Manifold-Constrained Hyper-Connections technology optimizes inter-layer information flow, surpassing the limitations of traditional residual connections.
Multimodality is another key differentiator. While V3 was text-only, V4 is designed as a unified multimodal model natively processing text, images, and video, signaling direct competition with GPT-4o and Gemini 2.0.
Price Disruption: One-Quarter of Competitors
DeepSeek's most potent weapon is undoubtedly pricing. V4's expected input token price is approximately $0.28 per million tokens, a dramatic figure compared to competitors.
| Model | Input Price | Notes |
|---|---|---|
| DeepSeek V4 (est.) | ~$0.28 | MoE 1T / 32B active |
| GPT-5.2 (OpenAI) | $1.75 | Latest frontier |
| Claude Opus 4.6 (Anthropic) | $5.00 | Standard 200K context |
| Gemini 3.1 Pro (Google) | $2.00 | Under 200K context |
| DeepSeek V3 (current) | ~$0.27 | MoE 671B / 37B active |
V4 maintains pricing nearly identical to V3 while dramatically improving performance. It's roughly 1/6th the cost of GPT-5.2, 1/18th of Claude Opus 4.6, and 1/7th of Gemini 3.1 Pro. This price competitiveness is attributed to lower labor costs in China and the efficient MoE architecture. If DeepSeek maintains its open-source policy, it becomes an attractive option not only for API users but also for companies seeking self-deployment.
Blackwell Chip Controversy: Self-Reliance or Export Control Violation?
The most contentious issue ahead of V4's launch concerns the chips used for training. DeepSeek claims training was conducted primarily on Huawei's Ascend 910B and Cambricon chips, positioning this as a case study in China's AI self-reliance. However, a very different narrative is emerging from the US side.
Multiple international outlets and industry analysts have raised the possibility that DeepSeek acquired a substantial number of NVIDIA's latest Blackwell B200 chips for training, allegedly obtained through circumvention routes despite US export controls on advanced semiconductors to China. Reports indicate that DeepSeek's GPU inventory appears significantly larger than officially confirmed.
This controversy transcends the corporate level, becoming a core issue in the US-China technology hegemony competition. If export control violations are confirmed, sanctions could affect not just DeepSeek but the entire Chinese AI ecosystem. Conversely, if training was genuinely accomplished with Huawei chips alone, it would mean China's semiconductor self-reliance is advancing faster than expected.
Conclusion: A New Phase in AI Hegemony
DeepSeek V4's arrival has the potential to fundamentally reshape the AI competitive landscape. With 1 trillion parameters, a 1-million-token context window, and dramatically lower pricing, if V4 delivers on its promised benchmarks, it will put considerable pressure on OpenAI, Google, and Anthropic.
The Two Sessions timing suggests this model is not merely a technological product but also a political message symbolizing China's AI capabilities. Depending on how the Blackwell chip controversy resolves, US export controls could tighten further, or China's self-reliance narrative could gain momentum.
On March 4, the world watches to see how DeepSeek's announcement will write the next chapter of the AI hegemony race.
