研 究
Research
重新定义语言模型的效率边界。
Technical Report· July 2026
Sparse Routing: How 3.2B Active Parameters Compete with 100B+ Models
We demonstrate that ultra-sparse mixture-of-experts architectures with carefully designed routing mechanisms can match or exceed the performance of dense models orders of magnitude larger.
阅读更多Blog Post· June 2026
CPU-First Inference: Optimizing Transformer Architectures for Consumer Hardware
A deep dive into the kernel optimizations, quantization strategies, and memory access patterns that make Fathom-1.0 run efficiently on commodity CPUs.
阅读更多Paper· May 2026
Expert Selection Dynamics in Ultra-Sparse MoE Networks
An analysis of expert selection patterns in 24-experts MoE layers. Routing converges to task-specific expert coalitions, enabling parameter-efficient specialization.
阅读更多Full Paper· July 2026
Fathom-1.0 Technical Report
Comprehensive technical report covering architecture decisions, training methodology, evaluation benchmarks, safety evaluations, and reproducibility details.
阅读更多