Masked diffusion needs more forward passes than uniform or Gaussian variants, the first proven separation in dLLM parallelism traced to critical window width.

SwarmLLM runs a 27B model across browser tabs via WebGPU and WebRTC. No install, 16.1 tok/s speculative decoding on a GB10, beating native llama.cpp decode.

LLM-42 from Microsoft Research uses verified speculation to make LLM inference reproducible across runs, with overhead proportional to deterministic traffic.

Wavel and MeshRT at SOSP 2026 show what AI compiler and runtime software must become when a chip exposes 900,000 cores on a single distributed mesh.

Z.ai’s GLM-5.3 built the inference stack for GLM-5.3-Flash on 100,000+ accelerators in 13 days, reaching 3.22x throughput through dense feedback engineering.

TypeSafe Jev outputs typed values with calibrated probabilities, not text. 70-500ms, $0.042/M tokens, schema-exact outputs for classification and routing.

A Cambridge hypernetwork generates live weight modulations from session data so LLMs adapt without frozen weights. arXiv:2609.18842, September 2026.

Nota AI cut GR00T N1.7 inference from 1,602ms to 230ms on a Qualcomm NPU, a 7x gain with no GPU, using custom runtime, graph rewriting, and step distillation.

DeepSeek R1 is a remarkable step forward in large language model training, known for its unique multi-stage reinforcement learning process and ability to create reasoning-focused AI. This blog dives into the training methodology of DeepSeek R1, demystifying its capabilities and showcasing its relevance in advancing AI technologies. What is DeepSeek R1? DeepSeek R1 represents a…