In the Smoke lightweight evaluation on July 1, 2026, Claude Opus 4.7 ranked first on the main leaderboard with a score of 94.82, while Gemini 3.1 Pro experienced a sharp drop of 32.2 points. The evaluation highlights a clear divergence between constraint and execution scores across models.
In the June 2026 Smoke evaluation of the YZ Index, Claude Sonnet 4.6 saw its main ranking score drop from 97.84 to 82.52 points, a single-day decline of 15.3 points, driven primarily by a 25-point fall in the code execution dimension.
In the YZ Index June 2026 Smoke Evaluation, Claude Opus 4.7's main score dropped from 100.00 yesterday to 84.01 today, and its code execution dimension fell from 100.00 to 72.80.
In the June 30, 2026 Smoke Lite evaluation of the YZ Index, Gemini 3.1 Pro ranked first with a main score of 98.47 points. Multiple models saw significant drops in execution scores, with Claude's execution scores plummeting over 25 points.
The security community has long relied on Coordinated Vulnerability Disclosure for handling dangerous findings, but this model fails for AI systems. MLCommons recognizes this as a core governance challenge for frontier model evaluation and is working to establish responsible disclosure standards.
The Chakra initiative by MLCommons aims to standardize AI system benchmarking through open execution traces, addressing fragmentation and enabling collaborative co-design across industry and academia.
MLCommons today announced the launch of MLPerf Mobile v6.0, adding generative AI benchmarks for running large language models (LLMs) on Android devices. These tests join existing benchmarks for image generation, object detection, and super resolution in the MLPerf Mobile app to form a complete test suite.
MLPerf Training v6.0 introduces two new Mixture-of-Experts benchmarks, DeepSeek V3 and GPT-OSS 20B, reflecting the industry's shift toward sparse computing. With 95 unique systems submitted by 24 organizations, the results show increasing diversity in hardware and software.
Optimizing GLM4-MoE for Production: 65% Faster TTFT with SGLangNovita AIJanuary 21, 2026TL;DR A suite of production-tested, high-impact optimizations has been developed by Novita AI for deploying…
Squeezing 1TB Model Rollout into a Single H200: INT4 QAT RL End-to-End PracticeSGLang RL Team, InfiXAI Team, Ant Group Asystem & AQ Infra Team, slime Team, RadixArk TeamJanuary 26, 2026 💡 TL;DR:…
No Token Left Behind: Demystifying Token-In-Token-Out in MilesMiles Team: Jiajun Li, Yuzhen Zhou, Shi Dong, Yanbin Jiang, Mao Cheng, Yusheng Su, Yueming Yuan, Zhichen Zeng, Banghua ZhuJune 5, 2026In…
Win on TCO: How AMD Instinct™ MI355X Achieves Cost-Competitive Distributed Inference Through SGLang with MoRIAMD & SGLang TeamMay 28, 2026The SGLang and AMD team has worked closely to unlock…
Heterogeneous CPU + GPU EPD Disaggregation to Boost VLM ServingIntel & SGLang TeamMay 29, 2026TL;DR We enabled heterogeneous Encode-Prefill-Decode (EPD) disaggregation via Dynamo and SGLang for…
SGLang and Miles Add Day-0 Support for NVIDIA Nemotron 3 Ultra for Long-Running Autonomous AgentsNVIDIA Nemotron Team and SGLang & Miles TeamJune 4, 2026 We are excited to announce that SGLang…
Higgs Audio v3 TTS on SGLang-Omni: Real-Time, Controllable Speech for Voice AgentsBoson AI & SGLang-Omni TeamJune 4, 2026Today we are announcing end-to-end serving for Higgs Audio v3 TTS on…
Announcing the Recipient of the 2026 LMSYS PhD FellowshipLMSYS BoardJune 08, 2026We are delighted to announce the first recipient of the LMSYS Fellowship Program: Will Lin. Following the launch of…
The next generation of speculative decoding: DFlash and Spec V2Z Lab, Modal, and SGLang TeamsJune 15, 2026Using Modal and Z Lab's DFlash speculative decoding models with SGLang’s newly default Spec…
MOSS-TTS Local Transformer v1.5 on SGLang-Omni: Serving Native-Streaming 48 kHz SpeechMOSI, OpenMOSS Team & SGLang-Omni TeamJune 17, 2026Today we are announcing end-to-end serving for…
Optimizing Ling-2.6-1T on TPU with SGLang-JAX: Hiding MoE Data Movement Behind Compute with One Pallas KernelPrayer, JamesBrianD, Haolin Fu, Haoguang Cai, Qinghan ChenJune 17, 2026SGLang-JAX now…
Improving DeepEP MoE Load Balance in SGLang with Waterfill and LPLBNVIDIA TeamJune 26, 2026TL;DR Mixture-of-Experts (MoE) models rely on Expert Parallelism (EP) to scale inference across multiple…
