A speech synthesis model with just 0.6B parameters that closes in on top-tier industry performance, Western-made semiconductors propping up Chinese AI companies, and the words "AI investment will eventually slow down" beginning to slip out of the mouths of major corporations. Here is a rundown of the notable moves from overseas AI news on July 31.
SOTA-Class Performance at Just 0.6B: Audio8 TTS Preview
"Audio8-TTS-Preview-0.6b", which trended on Hugging Face, is a remarkably small text-to-speech (TTS) model with only 0.6B parameters. Yet it offers zero-shot voice cloning and supports multilingual generation in 11 languages. According to the developer's public benchmarks, it delivers results on par with far larger models such as Fish S2 Pro and Higgs Audio v2, which have billions of parameters.
What makes it unique is the breadth of deployment options. It adopts a "DualAR" architecture that combines a slow AR layer with a fast one (inspired by Fish Audio S2 Pro), and with the ONNX INT4-quantized build it can reportedly run on CPU with roughly 1GiB of memory on an Apple M2 machine. For GPU environments, an SGLang Omni adapter is also available. It is fair to say this model pushes the recent drive to shrink and lighten high-quality voice AI to an extreme.
Moonshot Secures Around 20,000 Nvidia Chips from Alibaba
According to Bloomberg, Chinese AI company Moonshot has a compute agreement with Alibaba involving around 20,000 Nvidia-made chips. This hardware reportedly accounts for a substantial share of the compute Moonshot uses for its "Kimi" family of models.
The episode once again lays bare how much Chinese AI development still depends on Western-made semiconductors amid the US-China standoff. Earlier this month, the company unveiled "Kimi K3", an open-source model that comes close to the frontier models of OpenAI and Anthropic, instantly attracting worldwide attention. It suggests that behind these powerful models lies large-scale chip procurement that slipped through the net of export controls.
Profits Still Growing, but "It Will Slow Eventually": The Turning Point for AI Investment Reflected in Murata and Microsoft
Murata Manufacturing, which supplies AI components, announced strong results—including 81% growth in its data center business—and raised its full-year earnings forecast. At the earnings briefing, however, President Nakajima said the current pace of investment by data center operators will "slow down eventually," citing intensifying competition and rising debt levels.
This strangely echoes Microsoft's move reported the same day. The company signaled a slight trim to its massive data center investment plans, and Wall Street welcomed it, sending the stock higher. It may be a sign that investors and industry insiders are beginning to recognize the possibility that AI infrastructure investment—long assumed to keep climbing indefinitely—could enter a correction phase. That said, Murata itself still projects further growth in its data center business, so it is worth keeping in mind that this does not mean an immediate turn toward contraction.
LinkedIn Adds an "AI Slop" Report Button
LinkedIn has begun rolling out a "Seems like AI slop" button that lets users report low-quality AI-generated posts. It is a case of a platform starting to provide an actual feature to counter the flood of AI-generated content across social media.
While text and images have become easy to generate with AI, "AI slop" that amounts to noise for readers is becoming a social issue. The fact that a real-name, business-focused social network like LinkedIn has set up an explicit reporting path could well spread to other platforms.
Scanning Old Books, Then Discarding Them: The Controversy Over AI Training Data Procurement
Reports keep mounting that AI companies are buying up old books, scanning them, and then destroying them. The loss of physical books in the process of digitizing rare volumes has drawn criticism from the perspectives of copyright and cultural heritage preservation.
With securing training data now determining competitiveness in AI development, data collection in gray zones shows no sign of ending. At the same time, techniques for evading generative AI training and extraction—such as "Decoy Font (MixFont)", which makes text harder for AI to read—are also being proposed, and this contest looks set to grow ever more sophisticated.
DeepSeek-V4-Flash-0731 Follow-up: Local Adoption Spreads in the Community
*(Updated from our previous release report)*
"DeepSeek-V4-Flash-0731", officially released earlier today, keeps drawing wider reactions from the community. On Qiita, articles digging into the public benchmarks in detail have appeared, drawing renewed attention to the fact that "Flash outperformed Pro (preview) on every evaluation item." Meanwhile, in Reddit's LocalLLaMA community, unsloth has released a GGUF-quantized version, and moves toward running the model locally are advancing rapidly.
The simultaneous rise of the "Flash is actually the stronger one" upset and the demand for local execution looks like a development emblematic of open-weight models reaching maturity.
Wrap-Up
Models that pair smaller size with higher performance, data procurement operating in gray zones, and a bend in the road for the investment cycle. Today's news gave the impression that AI is shifting away from the one-way era of "as big as possible, as much as possible" and entering a stage where efficiency and sustainability are being called into question.