Notes by Jin

Models & Research Kimi K3.1 Pro: Advancing Open AI for Coding, 3D Design, and Financial Research — Kimi’s K3.1 Pro adds stronger coding, 3D design, and financial research capabilities, with a one-million-token context window. Kimi reports fewer factual errors and performance approaching leading proprietary models. It is available

Models & Research How OpenAI Is Strengthening Security Around Its Most Powerful AI Models — OpenAI tightened its training security after an internal AI model gained unauthorized internet access. It added stronger isolation, monitoring, and offline training environments, while using advanced AI agents to uncover weaknesses. Training resumed after fixes, with

Models & Research Title: Canva Introduces Craft, an AI Image Model Built for Design — Canva has launched Craft, its first internally developed image generation model, available in Canva and through a public API. Craft plans layouts and typography before rendering images at up to 2K resolution, aiming for greater accuracy

Models & Research Recursive Introduces Flint: A Fast Local Coding Model Built with Automated AI Research — Recursive’s Flint is an open-weight coding model that runs locally, keeping data private and avoiding API fees. It activates only 2.5 billion of its 40 billion parameters per token, delivering strong coding

Models & Research NVIDIA broadens its open model family for European languages and industries — NVIDIA is expanding its open model lineup across language, robotics, biology and quantum computing to help developers tailor applications to local languages and sectors. Additions include Nemotron 3.5 Super VL. Nvidia Blog ↗ My takeaway: Evaluate

Tools & Open Source Amazon launches TorchNeuron to loosen Nvidia's grip on PyTorch — The new developer tool aims to make it easier to run existing PyTorch applications on Amazon’s Trainium chips. Amazon says developers can switch with a single configuration change, without rewriting their code. The company

Models & Research Elorian Introduces Glimpse: AI That Reasons Through Images and Video — Elorian’s Glimpse revisits images and video to track objects, understand spatial relationships, and assess physical consequences. The company reports performance competitive with leading models at lower per-token prices, supporting applications in robotics planning, scientific analysis, and

Models & Research xAI launches Grok Voice 3.0 speech-to-speech model — xAI released Grok Voice 3.0, which ranked first on Artificial Analysis’s Speech-to-Speech Index with a score of 86.9. xAI reports response latency below one second and lower cost per task than comparable rivals in Artificial Analysis’

Models & Research Black Forest Labs Releases FLUX 3 Image Klein for AI Image Generation and Editing — FLUX 3 Image Klein is an open-weights AI model that generates and edits images using text prompts and up to 10 reference images. It supports text rendering and object placement controls, but requires

Models & Research Meta Releases Muse Spark Open, Bringing Advanced AI Agents to Local Hardware — Meta’s Muse Spark Open is an open-weight multimodal model built for coding, reasoning, and multistep tasks. It features 816 billion parameters and a one-million-token context window, and Meta says it can run locally on

Models & Research Jev 1.14: Steady Improvements Meet Experimental Capabilities — TypeSafe AI introduces Jev 1.14 with incremental performance improvements, especially in multilingual support, alongside experimental image inputs, a bounding-box output type, and longer context windows. These experimental capabilities give developers more room to explore, but the company cautions

Models & Research 💡 Gemini 4 Argon is officially now available for developers An agent builds and recursively improves Earth observation models — Researchers propose framing verifiable Earth observation (EO) tasks as a harness that defines the agent’s resources and a verifier that scores its models, letting an agent handle model

Models & Research Meta Introduces Muse Wonder: AI Agents Built to Remember and Follow Through — Meta’s new Muse Wonder and Muse Wonder Full models are designed for complex personal and professional tasks. Active Memory recalls relevant context across conversations, while Active Focus helps agents work through obstacles and sustain

Models & Research Statistical re-examination of METR’s AI time-horizon plot — Researchers recomputed METR’s 50% time horizons across 228 tasks and 26 AI systems using splines and item-response theory, relaxing the assumed relationship between human task time and AI difficulty. The work examines how accurately time horizons are estimated

Models & Research Microsoft introduces Decision-1, a model built for rapid structured decisions — Microsoft is positioning decision models as a new AI category distinct from LLMs, designed to return structured outputs that software can act on right away. Microsoft-Decision-1 is its entry in that space, available through Foundry. Microsoft ↗ My

Models & Research Qwen-Image-2.1-Turbo: AI Image Generation and Editing in Just Eight Steps — Qwen-Image-2.1-Turbo is a 7-billion-parameter AI model that generates and edits high-resolution images in eight steps. Its examples include portraits, typography, transparent images, and compositions using multiple reference images. HuggingFace ↗ My takeaway: Qwen-Image-2.1-Turbo is a

Models & Research Liquid AI Releases Fast, On-Device Decision Models — Liquid AI’s new open-weight models make structured decisions from text, images, or audio in a single processing pass. The company reports that d1-3B rivals a model 12 times its size, enabling fast local applications such as robotics and live

Models & Research Anthropic unveils Claude Haiku 5.5, its small, low-cost model — Anthropic describes Haiku 5.5 as its cheapest, fastest and most capable small model to date. It targets high-volume, cost-sensitive jobs such as summarization, compaction, database queries and classification. Anthropic ↗ My takeaway: According to Anthropic, Haiku 5.

Models & Research Mistral previews Large 4, a trillion-parameter multimodal model with open weights planned — Mistral has introduced a public API preview of Large 4, its largest model yet, targeting coding, agentic workflows, and enterprise tasks such as cybersecurity, finance, and engineering. The company reports competitive performance against leading open

Models & Research Reflection Unveils Beam, an Open-Weight AI Model Built for Efficient Coding and Reasoning — Reflection introduces Beam, a 501-billion-parameter AI model for coding, reasoning, and autonomous tool use, with 23 billion parameters active per token. The company reports competitive open-model performance while using three to four times less

Tools & Open Source Google pauses open source bug bounty amid flood of AI-generated reports — Google has paused its open source bug bounty program after a significant increase in AI-generated submissions, highlighting the pressure low-quality automated reports can put on human reviewers. TechCrunch AI ↗ My takeaway: Recently, I watched this

Models & Research Microsoft’s first streaming transcription model debuts atop Artificial Analysis accuracy rankings — Microsoft AI released MAI-Transcribe-2-Streaming, its first streaming speech-to-text model, which it says ranks first for both final and partial transcript accuracy on Artificial Analysis. The release also includes two text-to-speech models: MAI-Voice-2.1 and the