This is a series where you can "binge-read" the news you've accidentally missed, but are a little curious about, over the weekend. This time, let's check out the main news released mainly during the week of August 9th!
Sony and TSMC to establish joint venture to develop next-generation image sensors
Sony Semiconductor Solutions (hereinafter "Sony") and TSMC announced on August 11 that they have entered into a definitive agreement to establish a joint venture company for next-generation image sensors. The company name is Advanced Vision Semiconductor Manufacturing Co., Ltd., and the company is based in Koshi City, Kumamoto Prefecture, with mass production scheduled to begin in 2029.
The two companies reached a basic agreement on May 8th regarding a strategic alliance regarding next-generation image sensors, and this has now become a legally binding agreement.
Sony will invest approximately 465 billion yen and TSMC will invest approximately 282 billion yen, for a total of approximately 747 billion yen. Sony's investment includes cash and the transfer of assets through a company split. Sony will become the sole controlling shareholder, and the new company will be operated as a consolidated subsidiary of the Sony Group. It is said that the representative director will also be selected from Sony.
As for the division of roles, Sony will be responsible for the development of core technology for image sensors and product planning and design, while TSMC will provide cutting-edge process technology and manufacturing expertise. The company produces image sensors for smartphones that utilize cutting-edge manufacturing process technology.
Through this, the company aims to speed up product development based on customer needs, accelerate technological innovation, and strengthen supply capabilities.
Meta releases locally running 30B model “Muse Glimmer”
On August 10th, Meta released the open model "Muse Glimmer-30B" for AI agents. The license is Apache 2.0, and there are no restrictions on commercial use, modification, or redistribution.
Its unique feature is that it is designed to be run on your own PC. The model card is said to be ``created for autonomous agent processing on consumer hardware,'' and can be completed without using cloud equipment.
Quantization is the way to do that. The weights have been compressed to approximately 4-bit precision, and the language model portion has been compressed to less than 20GB. There are two types available: K-Quant-Dynamic for 32GB and K-Quant-17GB for 24GB of graphics memory (or integrated memory). The full-precision version requires 64 GB (same), and the 24 GB version requires approximately one-third of the required capacity.
What is worrisome is the drop in accuracy due to compression, but according to the company's measurements, the average drop in accuracy across 15 common benchmarks is only 0.2% for K-Quant-Dynamic and 1.0% for K-Quant-17GB.
It supports speculative decoding using the block diffusion model "DFlash" that predicts 16 tokens at once, and it is said that GeForce RTX 5090 has increased 3.1 times from 74.9 tokens per second to 233.4 tokens per second.
The Apple M4 Max has a 1.5x increase from 23.7 to 37.8 tokens, and the Apple M5 Max has a 1.8x increase from 26.6 to 50.2 tokens. For measurement, the RTX side uses llama.cpp and the Apple silicon side uses ExecuTorch.
As an execution environment, it can be used with compatible apps such as llama.cpp, Ollama, and LM Studio, and in addition to the base model, Hugging Face has assistant version, GGUF version, and ExecuTorch version.
The model itself has a total number of parameters of about 29.6 billion, of which about 1.8 billion are vision encoders. The input is text and images, the output is text, and the context length is over 131,072 tokens.
