🌟 Today's Headline
NVIDIA Cosmos 3: Unified Multimodal Foundation Model
NVIDIA has launched Cosmos 3, a groundbreaking unified foundation model integrating language, image, video, audio, and action modalities using a Mixture-of-Transformers architecture. Available in two sizes—Base Nano (16B with 8B reasoner + 8B generator towers) and Super (64B)—the model includes fine-tuned versions for Text2Image and Image2Video that now rank as the top open-weight models in their categories, nearly matching Nano Banana 2 performance. Cosmos 3 achieved leadership across 8+ open-model leaderboards in world reasoning and related benchmarks on launch day. The architecture combines an autoregressive reasoner with a diffusion generator, enabling sophisticated handling of multimodal reasoning and generation tasks. For developers and AI researchers, this represents a major step forward in accessible multimodal AI capabilities without relying solely on proprietary models.
💬 Editor's Note
When enterprise-grade open models hit SOTA performance, the real question shifts from capability to accessibility. Creators gain leverage over API pricing, but only if deployment friction—hardware cost, latency, optimization complexity—doesn't cancel out savings.
10/10
Tech
At Computex 2026, NVIDIA unveiled Nemotron 3 Ultra, a 550B-parameter open-weight large language model claiming the title of new US state-of-the-art (SOTA) among open models. The model is distinguished by remarkable efficiency and speed, delivering approximately 1.8x faster performance than competing open-weight alternatives on comparable tasks.
10/10
Industry
As AI agents become easier to use directly within companies, enterprise software vendors like Snowflake, Microsoft, and Databricks are fighting to prove they remain essential. Snowflake showed strength with a 33% stock jump and FY2027 product-revenue guidance raised to $5.84B. The company signed a five-year $6B AWS deal, with market cap reaching $90B.
9/10
New Product
datasette-agent-micropython 0.1a0 enables Datasette Agent to safely generate and execute Python code in a sandbox environment. Early testing shows GPT-5.5 and other frontier models have failed to escape the sandbox, suggesting the approach is promising for safe code execution workflows.
9/10
New Product
微软在Build 2026上发布了其首款高级推理AI模型MAI-Thinking-1。该模型被定位为"中等规模"，能在"关键"软件工程基准测试中达到领先模型的水平。微软称其完全从头使用干净数据进行训练，未涉及从第三方模型进行知识蒸馏。
9/10
News
Warren Buffett's Berkshire Hathaway invests $10 billion in Alphabet's AI infrastructure expansion. Alphabet raising $80 billion total with capital spending expected to reach $190 billion in 2026, signaling massive AI infrastructure buildout race.
9/10
New Product
在 COMPUTEX 上，NVIDIA 发布了 NemoClaw 平台，这是一个用于构建专业、长时间运行 AI 智能体的开放蓝图。该平台提供安全运行时、前沿模型支持以及多种编排框架集成选项，可通过 DGX Spark、数据中心或云端部署。
🕐 ~3 min read
· Opinion
6/10
💡 Views and arguments worth studying
This paper demonstrates that multiple weak preference signals from lower-quality model pairs (e.g., Qwen3 4B vs 1.7B) can be aggregated through LoRA merging to train strong LLMs. The relative quality deltas between weak responses provide effective supervision despite limitations in individual response quality.
🕐 ~4 min read
· Tutorial
6/10
💡 Can be adapted into tutorial material
This research introduces SemGrad, the first gradient-based uncertainty quantification method for free-form LLM generation. Unlike existing sampling-heavy approaches that are computationally expensive, SemGrad is sampling-free and computationally efficient. It leverages semantic-preserving embeddings to provide reliable uncertainty estimates, helping ensure LLM trustworthiness and reduce hallucination-related risks.
🕐 ~3 min read
· Tutorial
6/10
💡 Can be adapted into tutorial material
Zamba2-VL is a suite of vision-language models combining Mamba2 state-space layers with transformer blocks. It achieves competitive performance with leading open-weight VLMs like Molmo2, Qwen3-VL, and InternVL3.5 across image understanding, reasoning, OCR, grounding, and counting benchmarks.
Opinion
ThinkSwitch is a low-compute procedure for co-training paired instruct and thinking checkpoints to optimize inference-time reasoning. Starting from compatible Qwen3-4B models, the approach reduces latency and token costs while maintaining reasoning quality through efficient checkpoint interpolation.
A controlled experiment comparing 12 multi-agent LLM collaboration topologies for software architecture design using a 2×2×2 factorial design. The study ran 520 experimental runs across 8 design tasks with 5 repetitions each, evaluated by three independent automated evaluators including Claude Opus 4.6.
Audits Pali-to-English translation quality from GPT-5.5, Claude Sonnet 4.6, Gemini 3.1 Pro, and Grok 4.3 on 1,700 passages from the Pali Canon using three professional human reference translations. Addresses how single-score metrics conflate legitimate translation variation with actual translation errors.
Tutorial
在Code w/ Claude SF 2026活动上，Claude Code工程团队分享了将智能体编程设为默认工作方式后带来的流程与结构变革。核心变化包括：规划转向即时（JIT）模式，强调快速原型与反馈；上下文收集变为"先问Claude"；代码审查中Claude处理风格与测试，人工专注于法律、安全等…
MLLM-Microscope is a system for analyzing hidden representations in Multimodal LLMs by evaluating linearity, intrinsic dimension, and anisotropy of multimodal token embeddings across transformer layers. The study evaluates LLaVA-NeXT and OmniFusion on the ScienceQA dataset.
Introduces Prototype Transformer (ProtoT), an autoregressive language model that replaces quadratic-cost self-attention with a linear-cost prototype-based module, improving interpretability of LM reasoning and reducing hallucination risks.