China's Open-Source AI Blitz: Zhipu and Alibaba Drop Heavyweight Models on the Same Day

China's Open-Source AI Blitz: Zhipu and Alibaba Drop Heavyweight Models on the Same Day

GLM-5.3-Flash and Qwen3.8-Flash-Next go open-weight within hours of each other

In a striking display of competitive timing, two of China's biggest AI players released major open-weight models on August 26. Zhipu AI unveiled GLM-5.3-Flash, a 320-billion-parameter mixture-of-experts model that activates just 18 billion parameters per token. It's the first natively multimodal release in the GLM-5 series, supports a one-million-token context window, and comes with MIT-licensed weights on Hugging Face — priced at a tenth of Zhipu's flagship GLM-5.3 API rate.

Hours later, Alibaba's Qwen team dropped Qwen3.8-Flash-Next, a 125B-parameter model activating only 6B per token. The team explicitly framed it as a preview of the architecture behind Qwen4, with a native context of 262,144 tokens extensible to one million. Both models push the same design philosophy: massive total parameter counts with aggressive sparsity to cut inference costs without sacrificing capability.

The dual release underscores how Chinese labs are betting heavily on open-weight models to build ecosystem dominance — a strategy that's putting real competitive pressure on proprietary Western alternatives.

Nvidia's Groq 3 LPX enters full production — inference just got a lot faster

At Hot Chips 2026 on August 24, Nvidia announced that its Groq 3 LPX inference accelerator has entered full production. The chip is purpose-built for the Vera Rubin data center platform and represents Nvidia's dedicated answer to the ballooning demand for fast AI inference.

The numbers are eye-catching: a full LPX rack of 256 chips achieved 3,431 output tokens per second on a 100K-context Gemma 4 31B reasoning workload — roughly four times faster than the next-fastest public endpoint. Each LP30 chip carries approximately 500MB of on-die SRAM with no HBM, giving a full rack 128GB of memory at 40 PB/s aggregate bandwidth against 315 PFLOPS of FP8 compute.

Nvidia also disclosed the architecture of its Vera CPU at the same event: 88 custom Olympus cores across six chiplets on a single interposer, paired with LPDDR5X memory and NVLink-C2C for GPU connectivity. SpaceXAI has already committed to deploying Vera CPUs for the CPU-heavy workloads behind agentic AI — orchestration, tool use, code execution, and simulation.

Unitree's IPO rollercoaster sparks humanoid robotics bubble fears

The world's most closely watched robotics stock is having a rough week. Unitree Robotics, which became the first listed humanoid robot maker in mainland China after raising $904 million in its Shanghai IPO, saw its shares surge 460% on debut on August 18 before shedding roughly 45% of that value over the following sessions.

By August 24, Unitree closed at 603 yuan — a 40%+ drawdown from its opening-day high of 1,100 yuan — cutting its market cap to approximately $36.3 billion from a peak near $66 billion. Shares steadied on Tuesday but the damage has sparked broader concerns about whether the humanoid robotics sector is overheated.

The fundamental case is hard to make at current valuations: Unitree's adjusted net profit fell 53% to just $5.95 million in Q1 2026. Founder Wang Xingxing himself acknowledged at the World Robot Conference that humanoid robots “are not ready for broad factory deployment because they remain less efficient than humans at simple tasks and struggle to generalise across jobs.” The stock remains a barometer of investor appetite for the humanoid dream — and right now, that appetite is being tested.

Jeff Dean's departure reshapes Google's AI future

In one of the most significant talent moves the AI industry has seen, Jeff Dean left Google after 27 years to co-found Discovery Loop, a public benefit corporation focused on using AI to automate scientific and engineering research. He's joined by co-founders Sanjay Ghemawat, Oriol Vinyals (from DeepMind), and Quoc Le (from Google Brain) — a lineup that reads like a who's-who of modern AI research.

Discovery Loop initially focuses on automating machine-learning research before expanding into hardware design, drug discovery, and clean energy. Google itself is a founding investor and cloud partner, with Radical Ventures and Khosla Ventures co-leading the seed round.

The departure triggered a broader leadership reshuffle at Google: Demis Hassabis stepped down as Google DeepMind CEO to become chairman and Alphabet's chief scientist. The moves come at a pivotal moment for Google as it works to maintain its position against OpenAI and Anthropic in the intensifying AI race.

EU AI Act transparency rules are now live — and they have teeth

As of August 2, the European Commission's AI Office and national authorities began enforcing the EU AI Act's transparency obligations under Article 50. The rules represent the first continent-wide requirements for AI self-identification:

• Chatbots must tell users they're interacting with AI, not a human
• Deepfakes — AI-generated or edited images, videos, and audio — must be labelled
• AI-generated content must carry machine-readable marks for automated detection

The penalties for non-compliance are substantial: fines of up to €15 million or 3% of worldwide annual turnover, whichever is higher. The enforcement marks a significant milestone in AI governance and is being closely watched as a model for regulatory approaches worldwide. Companies operating in Europe now face a clear choice: build transparency into their AI products or risk serious financial consequences.