The HuggingFace Board Gives a Model 4.7 Days, and the Survivors Are Losing Ground
Ten captures of the HuggingFace trending board hold 64 models. The average lasts 4.7 of them, and five of the nine survivors lost attention anyway.
Tencent Still Hasn’t Shipped Hy4. It Shipped Qwen a 3.8 Times Speedup.
A five-layer speculative decoding drafter from Tencent lifts Qwen3.8-27B agent throughput 3.84 times without touching the weights.
Google’s Z4M Ships 168,000 GiB of Local SSD. The Ratio Is What You Buy.
Google's Z4M machine series is GA with 168,000 GiB of local SSD and 400 Gbps RDMA. The 1.4 to 1 ratio decides whether you can use any of it.
Fireworks Says the Token Price Is a Decoy. The Cache Hit Rate Sets the Bill.
A Fireworks VP says token price hides the real cost and cost per task is the number to run. Cache hit rate explains most of the claimed saving.
HuggingFace’s Board Went From Language to Geometry in Six Days
HuggingFace trending models flipped toward 3D and video inside six days. A 4B 3D coder sits at rank three and a new image-to-3d model at rank two.
HuggingFace’s Number One Model Has 0.56 Percent of Rank Twelve’s Downloads
HuggingFace trending put a model with 44,939 downloads at number one while rank twelve held 8,032,442. The board ranks momentum, not load.
The AI Buildout Just Found a Constraint No Grid Operator Controls
Voters settled 41 data center ballot measures and kept the industry out of all but one. The 25 megawatt ceilings sit far below an 8 GW Ohio campus.
The On-Prem AI Server Boom Is a Price Rise Wearing a Growth Rate
On-premises HPC server market revenue rose 19.9 percent while units grew 4.4. Most of that growth is component price, not capacity.
AI Suppliers Accelerated While the Buyers Hit a Delivery Wall
Every AI supplier beat, three of four hyperscalers burned cash, and the reason is capacity arriving late.