Adaptive-K routing: up to 52% compute savings on MoE models

Best GPUs for Local LLM

Buying guide: price/performance of GPUs for local inference.

ACAR: Adaptive Complexity Routing for Multi-Model Ensembles with Auditable Decision Traces

ACAR is a framework for orchestrating multiple models, using self-consistency variance to route tasks to configurations with one, two, or three models. Implemen

MCF-MOE: Multi-level context fusion delivers more coherent expert routing in MoE models

A new framework fuses cross-layer semantics with local token interactions to stabilize expert selection in Mixture-of-Experts models, boosting routing consisten

Plano: AI agent framework reaches 5000 stars on GitHub

Plano, an open-source framework for developing AI agents, has surpassed 5000 stars on GitHub. The project focuses on small LLMs for routing and orchestration, w

LLM Routing on Consumer GPUs: Ray Tracing Cores Accelerate MoE by 218x

Groundbreaking research has demonstrated how Ray Tracing Cores (RT Cores) on consumer GPUs, typically idle during LLM inference, can be repurposed to accelerate

China's GPU Four race to power large models, eye inference surge

Four Chinese companies are racing to provide high-performance GPUs for large language model (LLM) inference. The push is driven by increasing demand for on-prem

Qwen 3.8 27B: Q3_xxs quantization holds up on a 16GB RTX 4060 Ti

NVFP4 for Qwen3.8 27B: 6,250 tokens/s on RTX 5090

QwenMix-3.7: Seven Tokens, a Merge, and the Future of Local Checkpoints

Ling-3.0 releases six base checkpoints: two sizes, three stages, no post-training

QwenMix-3.7: merging Qwen 3.8 and 3.6 over seven tokens

Qwen3.8-27B: offline knowledge recall regresses compared to Qwen3.6