https://www.tella.tv/video/meta-s-spark-model-the-cost-of-your-data-hce3

Tokyo-based Sakana AI has released Fugu-Cyber, a model specialized in cyber defense. It claims performance on par with GPT-5.5-Cyber and Mythos. But what interests me is not the score. It is the company’s capabilities and its commercialization strategy. Here is a breakdown of the announcement, along with the key points that still need to be verified. ✏️ from choiopenai

Bloomberg reports that Z.AI, the developer of GLM, has completed a 1 GW data center filled entirely with Chinese-made chips and has begun partial operations. The facility has enough power capacity to supply roughly 750,000 households at once. It will be used to train next-generation GLM models, without a single NVIDIA chip. Some argue that fully populating 1 GW of capacity would require hundreds of thousands of Huawei Ascend chips, meaning that only part of the facility is likely operational today. But the direction is already being validated in practice. Z.AI says it trained its latest GLM models on Ascend chips using Huawei’s in-house software, completing the entire training pipeline on a domestic technology stack. Just two weeks ago, Kimi paused new subscriptions because it could not meet demand for K3 due to a shortage of GPUs. It was a clear reminder that the bottleneck for Chinese labs is not the models, but compute. China is now attempting to break through that export-control-driven bottleneck head-on by scaling up domestic chip capacity. from choiopenai

Alibaba’s Tongyi Lab has released Qwen-Audio-3.0-TTS. It comes in two versions: Flash for real-time use and Plus for higher-quality output. It supports 16 languages, allows users to control tone through natural-language instructions, and can even add nonverbal expressions such as laughter and breathing through tags. It ranked first on the independent Artificial Analysis TTS leaderboard. However, the top-tier version is not open-weight. It is available only through Alibaba Cloud’s API. Just days earlier, Alibaba had announced that its flagship LLM, Qwen3.8, would be released as open-weight. Yet its top-ranked voice model remains behind a paid API. Voice carries a higher deepfake risk and is also a lucrative real-time API market. That is likely why the company’s best-performing voice model will continue to remain behind an API. from choiopenai

A path has opened for fine-tuning LLMs on AMD graphics cards without NVIDIA. Unsloth has partnered with AMD to enable training and inference for more than 500 models across Radeon, Instinct, and Ryzen hardware. It works on Windows, WSL, and Linux. Unsloth says training can be up to twice as fast while using 70% less VRAM, with no loss in accuracy. Until now, fine-tuning has been effectively NVIDIA-only. Most performance optimizations were built on top of CUDA. Unsloth implemented those optimizations as Triton kernels, and because Triton is not tied to CUDA, they can run on AMD’s ROCm stack as well. This means even older cards with as little as 3 GB of VRAM can fine-tune models such as Qwen and Gemma. We will likely see much more local LLM training going forward. from choiopenai