Real testing, real data. We evaluate AI models, smart hardware, and cutting-edge tech with rigorous methodology — giving you the most objective reference.
The EU has selected MLPerf Training v5.1 and MLPerf Inference v5.1 as performance benchmarks for its AI Gigafactories tender, while MLCommons has opened submissions for MLPerf Endpoints v1.0 to evaluate real API-served AI services across performance, cost, and energy consumption.
MLCommons has updated its mission to focus on standardizing AI measurement and enabling informed decisions. It will continue building and operating industrial-grade benchmarks while evolving its organization and outreach.
MLPerf Endpoints evaluates inference systems as complete production endpoints under real workloads, presenting a Pareto front across concurrency levels and metrics directly tied to deployment decisions. The article explains what the benchmark measures, how it differs from other benchmarks, and how buyers should interpret its results.
MLCommons has been selected as the benchmark team for the STRIDE Ventures AI Efficiency Challenge and will use a two-year, $3.5 million NSF-supported grant to develop MLPerf Agents, a new benchmark for measuring the efficiency of agentic AI systems in completing real work.
SGLang Adds Day-0 Support for NVIDIA Nemotron 3.5 Super VLNVIDIA Nemotron TeamOctober 15, 2026Always-on AI agents must do more than generate text. They need to interpret visual information, reason acr
SGLang and Miles Add Day-0 Support for Muse Spark OpenSGLang TeamOctober 15, 2026Highlights Model: Muse Spark Open is an open-weight multimodal mixture-of-experts model. It uses an 816B-parameter text
SGLang Powers Beam in ProductionSGLang TeamOctober 16, 2026Reflection’s Beam is a 501B-parameter mixture-of-experts model with 23B active parameters per token, built for software engineering and long-
SGLang Low-Latency Mode: Kernels and RuntimeSGLang, NVIDIA and AMD teamsOctober 18, 2026 Low-latency inference offers many opportunities for optimizations tailored to specific models, hardware and wor
SGLang-TorchTPU: PyTorch-Native Production Serving on TPUsRadixArk, Google and MetaOctober 20, 2026 TL;DR SGLang-TorchTPU is open source today, built by the SGLang team at RadixArk in collaboration wi
The latest WDCD v3.1 pilot round shows GPT-6 Luna scoring 100.00 to top the current Top 5, while Doubao Pro jumped 20.3 points over Run #384 and Gemini 2.5 Pro and GPT-5.5 each dropped 6.2 points. Four models rose and two fell across the 15 models evaluated on a mixed pool of multi-turn progressive-pressure questions and three-round anchor questions.
In the WDCD v3.1 pilot, worst-of-3 scores for 15 models across five constraint scenarios show that engineering specification and data boundaries are the hardest dimensions for commitment compliance, with lows of 1.95 and 1.7, respectively.
In the WDCD v3.1 pilot, 15 models averaged a 0.93 R1 acknowledgment rate, a 0.67 R2 resistance rate, and only a 76.7% R3 integrity rate, with two full-collapse cases out of 150. The decay pattern shows that some models maintain constraints across all three anchor rounds, while others verbally agree but break under pressure.
