GOOG251128C260000.US

0715 | Dolphin Focus: 🐬 Stock 1, $TENCENT(00700.HK) rolled out Hy3 1-bit and 4-bit quantized builds. The original 295B model carried ~600GB of weights; the 1-bit build compresses this to 85.5GiB, allowing local runs on a single 96GB pro inference card.

The 4-bit variant delivers accuracy close to the full model. It is designed for enterprise offline inference.

Previously, large MoE models typically relied on multi-GPU clusters. The new quantization sharply lowers the hardware bar for smaller developers, strengthening Tencent’s on-device and private deployment capabilities...

In Dolphin Research's earlier piece 'AMZN's AI Misfire: Can It Mount a Comeback?', we examined AWS's AI buildout from several angles.

0708 | Dolphin Research Focus: Macro/Sector.