SenseTime’s ‘Big Device’ Powers the Leap of Chinese AI from Usable to Practical

A Delayed Adaptation Beats the Release Itself

DeepSeek V4’s release was slower than expected because the team chose to rewrite the software stack and perform low‑level optimizations for Huawei’s Ascend chips, targeting operator‑level and memory‑scheduling characteristics rather than merely getting the model to run.

This slowdown is intentional: tighter U.S. export controls make Nvidia’s high‑end AI chips harder to obtain, so domestic chips must fill the gap. However, hardware alone does not guarantee efficiency; the real gain comes from sophisticated compute scheduling. SenseTime’s “Big Device” assembles these domestic chips into usable clusters and provides end‑to‑end support from training to inference.

According to CXO Digitalpulse, Huawei confirmed that its Ascend infrastructure will fully support the new model, but current Ascend chips have throughput limits that will only be mitigated when the next‑generation chips become widely available later in the year. Consequently, SenseTime must shoulder the “software补全” (software completion) role during this hardware‑maturity ramp‑up.

Million‑Token Context: How SenseTime’s Big Device Provides the Back‑stop

Technically, DeepSeek V4 introduces several upgrades, most notably a massive 1 million‑token context window and a suite of architectural tweaks that lower compute and memory costs. Multiple model variants are offered to balance performance and cost for different enterprise and developer scenarios.

Such capabilities place extreme demands on the compute substrate: a million‑token window stresses VRAM, bandwidth, and distributed‑training communication efficiency. SenseTime’s support manifests in three dimensions: (1) fine‑grained scheduling of Ascend chip clusters, (2) mixed‑precision training and model‑parallel strategies, and (3) inference‑stage latency and cost reductions.

Key Takeaway

DeepSeek V4 is the first non‑Nvidia‑native model to achieve large‑scale deployment on Ascend chips. SenseTime’s Big Device acts as middleware—connecting the model upward and managing the hardware downward—making the soft‑hardware co‑design of domestic AI move from theory to practice.

Historically, domestic models on domestic chips were “good enough” but suffered efficiency penalties. With SenseTime’s assistance, the same model now runs on Ascend with performance that is closing the gap to Nvidia platforms.

The Domestic AI “Iron Triangle” Takes Shape

When DeepSeek V4, Huawei Ascend, and SenseTime’s Big Device are considered together, a clear domestic AI ecosystem loop emerges: DeepSeek supplies cutting‑edge model capability, Huawei provides autonomous, controllable compute hardware, and SenseTime bridges the final mile.

Industry analysts view DeepSeek and Huawei as core components of the Chinese AI ecosystem, with SenseTime’s value lying in making this core truly operational. Without an efficient compute platform, even the strongest models remain theoretical.

“This collaboration shows that domestic AI has moved from a ‘single‑point breakthrough’ to a ‘system‑level co‑ordination’ stage. The closed loop formed by SenseTime’s Big Device, DeepSeek, and Huawei Ascend finds a workable balance among performance, autonomy, and scalability.” – Industry observer

Challenges remain: current Ascend chip throughput limits mean large‑scale commercial deployment of DeepSeek V4 will require the next‑generation chips. Nonetheless, the direction is clear: domestic AI is no longer merely a follower but is building an independent full‑stack from chip to model to application.

When compute becomes the new oil, companies that control the compute foundation hold the lifeline of the AI era. The partnership of SenseTime, DeepSeek, and Huawei Ascend is not just a technical collaboration but a stress test for the commercialization of Chinese AI, with results expected in the second half of the year.