The highly anticipated DeepSeek V4 is coming soon! This month, Liang Wenfeng's team confirmed that it will be officially released in late April 2026. Here are several important feature signals of DeepSeek V4: one ️⃣ Trillion level parameters, million token context window, actually now supports input of millions of tokens; two ️⃣ The new Engram conditional memory architecture - Transformer's memory challenge, they say it has been solved; three ️⃣ Running Huawei Ascend chips, not relying on NVIDIA CUDA, completely domestic multimodal AI models; four ️⃣ Priced at $0.30 per million tokens, 10 times cheaper than GPT-5 and 16 times cheaper than Claude Opus 4.6, with excellent value for money! five ️⃣ Native multimodal, AI agents become standard, greatly improving mathematical reasoning and programming abilities; six ️⃣ Open source weight synchronization has been released, and the web version has launched a layered design of "Quick Mode" and "Expert Mode", which is being previewed for V4. The platform has experienced large-scale service anomalies in the past three days, most likely due to model switching. Trillion parameters+Huawei chips+1/10 price+multimodal, Liang Wenfeng is redefining the three words "cost-effectiveness". deepseek