Engineering
MobileMem: Learning from a Year of Mobile Experiences
A new benchmark called MobileMem evaluates AI agents’ ability to provide persistent personal assistance by synthesizing a full year of…
Engineering
Intern-S2-Mobius: Foundation Model with Decoupled Knowledge and Reasoning
Researchers introduced Mobius-v0, an architecture that separates knowledge storage from reasoning by using a globally shared memory implemented as a…
Engineering
DFM Mimir v1: An Open HRM Delivering Frontier Performance at 1B Parameters Using Only Permissible Post-Training Data
Researchers released Mimir v1, a 1-billion-parameter language model built on the Hierarchical Reasoning Model architecture and trained exclusively on ethically-sourced,…
Engineering
Scaling Domain Data Repetition in LLM Pretraining
A study examines how many times high-quality domain-specific data should be repeated during LLM pretraining as models and token budgets…
Engineering
Dion3: Full-Stack Orthogonal Updates
Dion3 targets the computational bottleneck of orthogonalizing weight updates in the Muon optimizer, introducing a Gram Newton-Schulz algorithm, specialized kernel…
Engineering
GLM-5.3: How Chinese Labs Keep Stride With the Frontier
Nathan Lambert’s Interconnects.ai analyzed how Z.ai’s GLM-5.3 model reaches frontier-level agentic coding performance without changing its roughly 750-billion-parameter base model…
Engineering
Distributed Layerwise Offload: Scaling Toward 200B+ DiT Models Efficiently in vLLM-Omni
The vLLM project detailed Distributed Layerwise Offload, a technique that lets large diffusion transformer models run across multiple GPUs or…
Engineering
Qwen 3.8 27B Is Excellent, But It Defaults to Wildly Overthinking Things
Simon Willison tested Alibaba’s Qwen 3.8 27B, a 27-billion-parameter vision-capable open-weight model that fits in a 17GB file, and found…
Engineering
Adaptive Verification in vLLM: DSpark confidence-scheduled verification
vLLM implements adaptive verification in speculative decoding using DSpark’s confidence-scored drafting mechanism, which dynamically adjusts verification length per step instead…
Engineering
Day 0 Support for Qwen3.8-2.4T-A95B on vLLM
vLLM announced Day 0 support for Qwen3.8-2.4T-A95B, a 2.4-trillion-parameter sparse mixture-of-experts model with 512 experts, running without architecture modifications. The…
Engineering
vLLM Reaches 25K Total TPS/GPU on Qwen3.5
vLLM reached 25,000 tokens per second per GPU serving Qwen3.5 on Blackwell GPUs via a disaggregated prefill-decode architecture. Three optimizations…
Engineering
LEMUR and Mean Centering for Late-Interaction Retrieval in txtai
txtai adds LEMUR and mean-centering techniques for late-interaction retrieval. LEMUR learns a fixed-dimensional encoding of multi-vector representations so late-interaction models…