MLCommons' MLPerf Inference working group has launched the first end-to-end retrieval-augmented generation (RAG) inference benchmark, measuring the full pipeline from document ingestion to multi-hop question answering across two workloads.
This article explores how double-blind evaluation design protects privacy and benchmark integrity, highlighting MLCommons' first proof of concept for closed-source model evaluation and its path toward becoming an industry standard.
Neocloud Lambda has raised $1B in private debt to buy Nvidia AI chips and lease them to Microsoft. It's the latest in a string of loans, underscoring the high cost of the AI boom.
Alibaba's Qwen team has released and open-sourced Qwen3.8-Flash-Next, a 125B-parameter multimodal MoE model with only 6B active parameters per token, priced at approximately 3% of Claude Opus4.6 with training costs reduced by about 90%. Its "Next" architecture is positioned as the technical prototype for the upcoming Qwen4 flagship.
In August 2026, a federal judge in California struck down the Pentagon's blacklisting of AI company Anthropic, ruling that Defense Secretary Pete Hegseth's actions violated the First Amendment and the Fifth Amendment's due process clause. The landmark decision establishes that national security designations cannot be used to punish companies for criticizing government positions.
Anthropic refused to support lethal autonomous warfare and mass surveillance.
There's a lot of capital pouring into the business of giving models away.
Given 10 benchmarks for specific misaligned behaviors, the automated systems were able to improve performance on every single one without degrading overall performance.
In today's Smoke evaluation, Claude Sonnet 4.6's code execution score plunged 22 points to 75.00, while its material constraint score surged 25.7 points to 93.30; the main leaderboard score slipped just 0.5 points to 83.24. The extreme inverse swings are attributed to small-sample question sampling volatility rather than genuine model degradation.
Claude Opus 4.7 scored 83.24 points on today's Smoke evaluation main leaderboard, down 10.3 points from yesterday's 93.54, driven by a drop in the code execution dimension from 97.00 to 75.00 points.
The YZ Index Smoke quick test on 2026-08-29 covered 11 models. Grok 4 topped the day with 96.99 points, while notable gains were seen for Gemini 3.1 Pro, GLM-4.6, DeepSeek V4 Pro, and Grok 4, and sharp declines were recorded for GPT-5.5 and several Claude models.
Pushing the Limits of Serving DeepSeek-V4-ProTianyu Zhang, Yusong Gao, Yun ZhangAugust 19, 20261. Introduction DeepSeek-V4-Pro is a 1.6-trillion-parameter Mixture-of-Experts (MoE) model released with
Mooncake for Miles: From Fragmented Rollout Data to Efficient Bulk I/OMooncake communityAugust 20, 2026Rollout Data in Disaggregated RL Systems Reinforcement learning for large language models combine
Fast Engine Recovery: Sub-Second Engine Restart for SGLang via Weight Cache DaemonAnt Ling Infra Team (Ant Group), Alibaba, SGLang TeamAugust 21, 2026TL;DR Nowadays, State-of-the-Art (SOTA) models are
Chasing the Batch-1 Floor: Ling-3.0-flash Speculative Decode on BlackwellRadixArk SGLang Team, Ant Ling Infra TeamAugust 21, 2026Batch-1 decode keeps getting more important. Xiaomi MiMo, for example,
Qwen3.8-Flash-Next: Day-0 Support in SGLangSGLang TeamAugust 26, 2026Introduction Today, the Qwen team open-sourced Qwen3.8-Flash-Next, a multimodal MoE model and an early preview of the Qwen4 archite
MiniMax-H3 on 8×H200: 1.95× Lossless, Up to 6.24× at 0.76–0.91 SSIMSGLang Diffusion Team, Cache-DiT Team, NVIDIA, Ant GroupAugust 27, 2026TL;DR We benchmarked MiniMax-H3 video generation on 8× NVIDIA
Infer-forge: Harness, Loop, and Graph Engineering Around SGLangTianyu Zhang, Hanlin Gao, Yusong Gao, Yun ZhangAugust 28, 20261. Introduction Inference optimization may look local in code, but its vali
A recent paper argues that AI is often better at doctoring than doctors. Guess who isn't thrilled.
Meta fixes AI glasses to stop recording any time users cover up the safety light.
