Thinking Machines Lab Releases 975B-Parameter Model "Inkling"

Thinking Machines Lab, founded by former OpenAI CTO Mira Murati, has released Inkling, a 975-billion-parameter multimodal open-weight model. It posted the highest score of any American open-weight model on the Artificial Analysis Intelligence Index, though it still trails China's top models on some tasks.

Inkling's input token pricing starts at $1.87 per million tokens. The company says it positions Inkling not as "the most powerful model" but as a base for fine-tuning. As competition among open-weight models heats up — and with Murati's name recognition — its next moves will be worth watching.

Apple May Partner with PrismML — Shrinking AI to 1/15 Its Size on iPhone

According to CNBC, Apple is exploring a partnership with PrismML, a company with technology that can compress AI models by up to a factor of 15. The technology makes it possible to run large models efficiently on devices like the iPhone, and could mark an important step toward practical on-device AI.

Interest is growing in on-device AI, which protects privacy without sending data to the cloud, and Apple could be gearing up for a major move in this space.

TSMC Raises Revenue and Investment Outlook on AI Demand

TSMC, the world's largest contract chipmaker, has raised its revenue and capital expenditure forecasts on the back of sustained AI-related demand. Reported by Bloomberg, the news suggests that momentum in AI infrastructure investment remains strong.

Markets hold a range of views on how durable the AI boom is, but TSMC's decision to guide higher on both revenue and investment looks like further evidence that the "AI megatrend" is underpinned by real demand.

Google "Quietly" Updates Gemma 4 — Fixes Shipped Under the Same Name

Google has shipped an update to its open AI model, Gemma 4, without changing the version name. The update reportedly includes performance improvements on Nvidia Hopper GPUs, bug fixes for tool calling, and a fix for responses getting cut off mid-stream.

The improvements themselves are welcome, but shipping updates under the same name — a "silent update" — leaves concerns from a reproducibility and evaluation standpoint. Calls for transparency are likely to come from the developer community.

Japanese AI Developer "Noetra" Launches in Earnest — With Help from NVIDIA

Noetra, an AI development company jointly funded by major Japanese corporations, has officially begun work on developing a homegrown Japanese AI model. With the cooperation of NVIDIA, the company plans to build a computing infrastructure equipped with 27,500 units of the next-generation "Rubin" GPU architecture.

Building Japan's own foundational AI technology is also significant from an economic security perspective. What kind of models it will produce — and how things unfold from here — will be closely watched.

No New Hires for Jobs AI Can Do — Policy at a Major Australian Mining Company

A major Australian resources company has reportedly told employees that it will not make new hires for work that AI can handle. It's a concrete example of AI's impact on employment beginning to surface as explicit corporate policy.

Many companies are pursuing AI-driven efficiency, but with it now manifesting as hiring freezes, the broader societal impact has been thrown into sharp relief once again.

---