xenospectrum

Fractile Raises $720 Million to Build AI Inference Chips That Pair Memory Speed With Capacity

UK-based Fractile has raised $720 million, with two UK government-backed bodies joining the round. Here is what its stacked ...

[2026 Latest] llama.cpp MoE cache is off by default. A trap where it doubles speed if VRAM is low, but slows things down if you have plenty

For those of you who tweeted last week, 'I have 24GB installed, but big models still won't run,' there is something I want ...
eu.36kr

Build the next-generation LLM inference engine oriented to Agentic reasoning

This in-depth guide systematically unveils the full technical roadmap of TokenSpeed, spanning every critical phase from ...