DeepSeek V4 Next Week? New Paper Reveals DualPath Inference Framework

The DeepSeek team has once again demonstrated its ambition for "system-level" evolution. While developers worldwide are scouring the web for the DeepSeek V4 release date, a paper titled "DualPath," co-published by the DeepSeek system group with Peking University and Tsinghua University, has quietly appeared on ArXiv. This is more than an academic release; it’s a preview of the core inference architecture of DeepSeek V4.

The DualPath framework directly addresses the most painful bottleneck in AI Agent deployment: I/O constraints during long-context inference.

Breaking PD Separation: The Art of Using "Idle Bandwidth"

In traditional Prefill-Decode (PD) disaggregated architectures, all KV-Cache loading tasks are crowded onto the storage network cards of the prefill engine, causing instant bandwidth saturation. Meanwhile, the network cards of the decoding engine remain idle.

DualPath introduces a second "Storage-to-Decode" path, enabling global pooling of cluster storage bandwidth. In real-world tests with 660B scale models:

• Offline inference throughput increased by 1.87x.
• Online service throughput increased by an average of 1.96x.
• Time to First Token (TTFT) was significantly optimized, making Agent responses feel almost instantaneous.

The Final Piece for 24/7 AI Employees

Why go to such lengths to optimize KV-Cache "transport" speed? The answer lies in the 24/7 autonomous digital employee scenario.

Future AI agents won't just be "Q&A" chat boxes. As demonstrated by OpenClaw, they will be capable of executing tasks 24/7, reading thousands of emails, and managing complex decision chains. In these scenarios, KV-Cache hit rates often exceed 95%. The DualPath framework makes running long-context Agents cheaper and faster by accelerating the retrieval of "old memories."

This means that when DeepSeek V4 is deployed within a localized Agent framework like OpenClaw, your AI employees will have nearly infinite instantaneous memory without the "thinking lag" caused by heavy workloads.

DeepSeek V4 Release Date Speculation

While official dates haven't been confirmed, the lead author of the DualPath paper, Yongtong Wu, is currently a member of the DeepSeek system group. Combined with recent CNBC warnings about the market impact of DeepSeek's new model, industry consensus is building that DeepSeek V4 is highly likely to debut as early as next week.

This DualPath architecture could be the "secret weapon" that allows V4 to surpass GPT-5 in Agent performance while maintaining extreme cost-efficiency.

👉 Don't want to miss next week's big reveal? Visit DeepSeekV4.app to subscribe to our real-time insider reports and get the scoop before the news hits Twitter.

---

Editor: UIO