It was supposed to launch during the Lunar New Year celebrations. Then early March. Then March 9. Each deadline passed with nothing but silence from Hangzhou.
DeepSeek V4, China’s answer to GPT-5 and Claude Opus, remains unreleased after five missed launch windows. The trillion-parameter model that was meant to demonstrate China’s AI independence from Nvidia hardware has become the industry’s most elaborate disappearing act.
What We Know
The specifications, leaked through Financial Times reporting and private beta testers, are ambitious. According to NxCode’s technical analysis:
- Approximately 1 trillion total parameters in a Mixture-of-Experts architecture
- Only 32-37 billion parameters activate per token, keeping inference costs manageable
- 1 million token context window, up from V3’s 128K
- Native multimodal capabilities for text, image, and video generation
- Optimized for Huawei Ascend and Cambricon chips, not Nvidia
The benchmark claims are aggressive. Leaked scores suggest 90% on HumanEval and over 80% on SWE-bench Verified, matching or exceeding Claude Opus 4.5. None of this has been independently verified.
The March 9 Mystery
Something did happen on March 9. Chinese tech outlet Sina Tech reported that DeepSeek’s website showed a silent update with “improved coding performance and expanded context handling” visible to testers.
The community calls it “V4 Lite” or by its internal codename “Sealion Lite.” Leaked specifications suggest roughly 200 billion parameters with the same million-token context window as the full V4. Inference provider PixVerse confirmed private testing of the 1M context capability.
DeepSeek published nothing. No blog post, no API documentation, no acknowledgment that anything changed. The market for “Will DeepSeek release V4 in March 2026?” now sits at 26%, down significantly as days pass without an announcement.
Why The Delay?
Three explanations have emerged, each revealing something about the state of Chinese AI development.
Hardware adaptation takes longer than expected. DeepSeek V4 might be the first trillion-parameter model to achieve full optimization for Huawei Ascend chips rather than Nvidia GPUs. Technical analysts suggest this isn’t just swapping drivers - it requires deep refactoring of operator libraries, communication topologies, and parallel training strategies.
This matters. If DeepSeek can prove frontier AI models train effectively on Chinese silicon, it undermines the entire premise of US chip export controls. But the engineering challenge is real. CUDA has decades of optimization. Huawei’s ecosystem is catching up, not caught up.
Multimodal integration is harder than expected. DeepSeek’s prior models excelled at text and code but lacked native image and video capabilities. V4 was meant to close that gap with multimodal training from the start. Integrating three modalities at trillion-parameter scale may simply take more tuning than anticipated.
Strategic timing. DeepSeek characteristically provides zero official communication. The silence may be deliberate positioning - waiting for the right geopolitical moment to demonstrate capability. March 5’s parliamentary discussions around AI self-reliance would have been one such moment. It passed without a launch.
The Dual Release Theory
Some observers expect DeepSeek to release V4 alongside R2, its next-generation reasoning model. The technical analysis notes that integrating two frontier models into a single, stable API requires extensive testing. If this theory holds, the delay reflects ambition rather than failure.
What This Means For Everyone Else
DeepSeek V4’s eventual arrival matters beyond China. The expected pricing - roughly $0.14 per million input tokens, half of V3’s cost - would pressure every commercial AI provider. The Apache 2.0 license means weights will be freely available for local deployment.
For the local AI community, this represents significant opportunity. If V4’s quantized versions run on consumer hardware as efficiently as V3’s derivatives, million-token context becomes accessible to anyone with a decent GPU. The privacy and cost implications are substantial.
But every week of delay gives competitors time to respond. OpenAI’s GPT-5 is shipping. Anthropic’s Claude Opus 4.5 has been available since January. Google’s Gemini 2.5 Pro dominates the long-context benchmark. Each passing day diminishes V4’s window to define the next generation.
The Bottom Line
DeepSeek V4 represents China’s most ambitious attempt to match Western frontier AI capabilities on domestic hardware. The silence around its delayed release suggests either significant technical challenges or calculated positioning. Either way, the trillion-parameter model that was supposed to reshape the competitive landscape remains, for now, vaporware.
When it does ship - and betting markets still give better-than-even odds for March or April - we’ll finally know whether DeepSeek’s gamble on Huawei silicon paid off. Until then, we wait.