China's AI and technology sectors are moving in lockstep on multiple fronts. In a single day, DeepSeek launched its most capable open-source model yet, Huawei demonstrated same-day chip compatibility, Tencent unveiled a new large model pivoting toward monetization, and data confirmed that China's tech giants have poured billions into robotics. Taken together, April 24 offers a snapshot of a domestic AI stack that is increasingly self-contained — and increasingly competitive.

• DeepSeek Raises the Bar With Million-Token Open-Source Model

DeepSeek released preview versions of V4-Pro and V4-Flash on April 24, introducing a one-million-token context window — a 10x expansion over its predecessor — alongside a novel sparse attention architecture that significantly reduces memory and compute overhead at extended contexts. Both variants are immediately accessible via API and open-sourced on Hugging Face and ModelScope. A three-month deprecation clock has started for legacy endpoints, which will shut down July 24.

Why it matters: The V4 launch reinforces DeepSeek's role as the bellwether of China's open-source AI movement. Its January V3 release rattled global markets by demonstrating frontier-level efficiency at lower cost; V4 doubles down on that narrative with architectural innovation rather than brute-force scaling. The million-token window opens the door to document-heavy enterprise use cases — legal, financial, and software development workflows — where retrieval-augmented generation has historically struggled. The open-source release also continues to complicate U.S. export control strategies by distributing cutting-edge capabilities globally at no cost.

• Huawei Ascend Matches DeepSeek V4 on Day Zero

Within hours of DeepSeek's V4 release, Huawei announced full compatibility across its Ascend 950 and Atlas 900 A3 SuperPod lineup. The Ascend 950 achieves inference latency as low as 10 milliseconds for V4-Flash and processes 4,700 tokens per second on single cards. Huawei also introduced PyPTO, a Python-first programming framework designed to lower the barrier to developing custom AI operators on Ascend hardware.

Why it matters: Since U.S. export controls cut off access to Nvidia H100 and A100 GPUs in late 2023, Chinese AI developers have faced a stark choice between rationing legacy chips or migrating to unproven alternatives. Huawei's ability to deliver competitive inference metrics on launch day — and to do so using domestically mature NAND flash rather than import-constrained high-bandwidth memory — signals that the Ascend roadmap is maturing faster than many observers expected. The PyPTO tooling is a direct challenge to Nvidia's CUDA ecosystem lock-in, though enterprise adoption will take time. The remaining open question is training at scale: Huawei has yet to publish benchmark results for training workloads comparable to DeepSeek's 671-billion-parameter V3.

• Tencent's Hy3 Bets on Monetization Over Benchmarks

Tencent launched a preview of Hy3, a 295-billion-parameter mixture-of-experts model developed in under three months under AI architect Yao Shunyu. The model is already live across Tencent Cloud, QQ, Tencent Docs, and the WorkBuddy enterprise assistant, with integrations for popular open-source agent frameworks including OpenClaw and KiloCode available via its TokenHub platform.

Why it matters: Hy3's rapid deployment reflects a broader industry shift in China's AI race — from chasing benchmark supremacy to capturing enterprise IT budgets and consumer subscriptions. Tencent's strategy is distinctive: rather than competing on raw capability, it is embedding AI directly into its existing ecosystem of gaming, social, and productivity platforms, using real-world user data as its primary evaluation loop. The model enters a crowded window alongside Alibaba's Qwen 3.6 Max Preview, Moonshot AI's Kimi 2.6, and Xiaomi's MiMo-V2.5-Pro — a sign that China's "second half" of the AI cycle is intensely contested.

• China's Tech Giants Deploy 2–2–5 Billion Into Robotics

Eight major Chinese tech companies executed 62 robotics investments in 2025, deploying between RMB 14.5 billion and RMB 34 billion (2–2–4.7 billion), according to IT Juzi data. The sector attracted RMB 73.5 billion ($10.2 billion) across 744 total deals. Meituan leads by deal count with 20+ investments, reporting 75–90% cost savings versus human couriers in drone delivery. Multiple giants — Alibaba, ByteDance, Meituan, and Xiaomi — simultaneously co-invested in autonomous robot startup Zibianling's RMB 2 billion Series B in April 2026.

Why it matters: The shift from passive investment to active hardware development marks a structural change in how China's internet giants view robotics — not as speculative R&D, but as operational infrastructure. The co-investment model, where rivals back the same startups, reflects robotics' unique economics: unlike digital platforms, physical robots require manufacturing scale and supply chain depth that no single company controls. Honor's humanoid robot winning a Beijing half-marathon seven minutes faster than the human world record triggered immediate supplier stock rallies — a market signal that the investment community now views supply chain readiness, not software demos, as the key variable.

What to Watch Next

The critical near-term question is whether Huawei Ascend can validate its training capabilities at the scale required for next-generation models. DeepSeek's full V4 release — and its pricing structure post-preview — will test how aggressively China's AI-as-a-service market compresses margins. On robotics, the race shifts from funding announcements to commercial deployment metrics: unit economics, not unicorn valuations, will determine which platform strategies survive.