Blog
Research notes, engineering deep-dives, and company updates.
Fathom-1.0: Intelligence, Distilled.
Today we release Fathom-1.0, our first production model: 3.2B active parameters, competitive with models 10–50× its size, running on consumer CPUs.
CPU-First Inference: Why We Optimized for the Hardware You Already Own
A technical deep-dive into the kernel optimizations, quantization strategies, and memory access patterns that make Fathom-1.0 run efficiently on commodity CPUs.
Sparse MoE at 24 Experts: Lessons from Training Ultra-Sparse Models
We share insights from training a 24-expert sparse mixture-of-experts model where only 3 experts are active per token, and what we learned about routing dynamics.
Our Responsible Scaling Policy: A Framework for Safe Capability Development
Wyrmlabs publishes its tiered Responsible Scaling Policy, modeled on ASL standards, with capability thresholds that trigger automated safety measures.
Benchmarking Fathom: Methodology, Reproducibility, and Results
A detailed look at our evaluation framework, including benchmark selection, protocol standardization, and our commitment to reproducible AI research.
Building an AI Lab in Beijing: Why Zhongguancun Is the Right Place
Reflections on building a world-class AI research lab in Beijing's premier technology district, and what it means for the global AI ecosystem.
Subscribe to our newsletter for the latest updates. contact@wyrmlabs.be