Apr 23, 2026

Key Highlights

• Qwen3.6-27B: Flagship-Level Coding in a 27B Dense Model

  • Qwen is a versatile AI platform, with multiple domains such as qwen.ai and chat.qwenlm.ai, indicating its broad accessibility and integration capabilities.

• Our eighth generation TPUs: two chips for the agentic era

  • Google’s eighth generation TPUs, TPU 8t and TPU 8i, are engineered for the agentic era, enhancing AI capabilities with specialized architectures for training and inference.

• Approximating Hyperbolic Tangent

  • Fast approximations of the hyperbolic tangent function (tanh) are essential in applications like neural networks and audio processing, where speed and efficiency are critical due to high-frequency evaluations and real-time performance demands.

• We benchmarked 18 LLMs on OCR (7k+ calls) — cheaper/old models oftentimes win. Full dataset + framework open-sourced.

  • Benchmarking revealed that smaller and older LLMs often outperform premium models in OCR tasks, demonstrating that cost-effective solutions can achieve comparable accuracy with less financial burden.

• GPT-5.5: Mythos-Like Hacking, Open to All

  • GPT-5.5 significantly enhances vulnerability detection, achieving a miss rate of just 10%, compared to 40% for GPT-5 and 18% for Opus 4.6, marking a substantial leap in offensive security capabilities.

• I can't believe text normalization is so underdiscussed in streaming text-to-speech

  • Text normalization in streaming TTS models is critically overlooked, leading to significant errors in pronouncing essential elements like dates, URLs, and promo codes.

• INT3 compression+fused metal kernels

  • INT3 compression achieves a notable +0.14 nats efficiency, enabling the development of a 2-bit KV cache tailored for long-horizon tasks, with the Qwen 7B model currently in preview.

• OpenAI’s New GPT-5.5 Powers Codex on NVIDIA Infrastructure

  • OpenAI’s GPT-5.5 powers Codex on NVIDIA infrastructure, enabling transformative results for over 10,000 NVIDIANs across various functions, significantly enhancing productivity and innovation in knowledge work.

• AutoAdapt: Automated domain adaptation for large language models

  • AutoAdapt automates the domain adaptation process for large language models (LLMs), transforming weeks of manual tuning into a repeatable, efficient pipeline that meets specific constraints like accuracy and latency. This framework utilizes a structured configuration graph and a budget-aware optimization loop to streamline the adaptation process, ensuring models are reliable in high-stakes environments such as law and medicine.

• Hybrid Policy Distillation for LLMs

  • Hybrid Policy Distillation (HPD) enhances knowledge distillation by merging forward and reverse KL methods, optimizing both mode coverage and mode-seeking for better model performance.

• Knowledge Capsules: Structured Nonparametric Memory Units for LLMs

  • Knowledge Capsules are structured nonparametric memory units that enhance large language models (LLMs) by allowing direct integration of external knowledge without retraining, thus improving efficiency and adaptability.

• V-tableR1: Process-Supervised Multimodal Table Reasoning with Critic-Guided Policy Optimization

  • V-tableR1 introduces a process-supervised reinforcement learning framework that enhances multimodal large language models (MLLMs) by enabling rigorous, verifiable reasoning rather than mere pattern matching.