vLLM 0.17 Ships FlashAttention 4 and Live MoE Scaling
vLLM v0.17.0 adds FlashAttention 4, elastic expert parallelism for live MoE rescaling, full Qwen3.5 support, and a performance-mode flag, all in 699 commits from 272 contributors.
Inference
AGI in 2026? Andrew Ng Has Some Bad News for Musk
Andrew Ng says AGI is decades away and the real AI bubble risk is in the training layer - not inference. We examine both claims against the data.
Mercury 2 Review: 1,000 Tokens per Second, Tested
Mercury 2 by Inception Labs is the fastest reasoning LLM available, built on diffusion architecture. We tested the speed, quality, and real-world trade-offs.
Mercury 2 Is 13x Faster Than Claude Haiku - Verified
Inception Labs' Mercury 2 hits 1,196 tokens per second in independent testing - a diffusion architecture that rewires how inference works.
NVIDIA's Secret Chip Fuses GPU and Groq for OpenAI
NVIDIA will unveil a new inference processor built on Groq's LPU architecture at GTC 2026, with OpenAI as its first major customer allocating 3 GW of dedicated capacity.
Someone Reverse-Engineered Apple's Neural Engine and Trained a Model on It
A developer cracked Apple's undocumented ANE private APIs, measured its real throughput at 19 TFLOPS FP16 (not the marketed 38 TOPS), and trained a 109M-parameter transformer on hardware Apple designed exclusively for inference.
