Welcome to Tech Beat, your daily look at the stories shaping our digital world.
The artificial intelligence community is getting a clearer picture of how leading mixture-of-experts models actually work under the hood. A new comparison across ten major models, including DeepSeek, Qwen, Mixtral, and Llama four, reveals striking disagreements on routing design. Expert counts range from eight to three hundred eighty four, and top-k selection varies from one to eight, meaning two models built on the same core idea can behave in fundamentally different ways at inference time.
Shifting to consumer hardware, Apple's MacBook Pro with the M five chip is now on sale, and one reviewer argues it may be the smarter buy over the freshly announced M six model. The performance gap between generations is narrowing, and for most users the M five already handles demanding workloads with room to spare, raising a familiar question about whether chasing the latest silicon is worth the premium.
And on a quieter corner of the maker internet, a developer has posted a technique for speeding up Bluetooth over-the-air firmware updates on the ESP thirty two microcontroller. It is a small story, but a telling one about the unglamorous, patient work that keeps embedded systems reliable in everything from smart home devices to industrial sensors.
Keep surfing. Tech Beat out.
