Muse Spark Open is an 816-billion-parameter (42B active) natively multimodal model with a 1,000,000‑token context window, built for long‑horizon coding, agentic workflows, and multi‑step reasoning. It is engineered to run on consumer hardware by pairing two Apple M5 Ultra Mac Studios via the MLX stack, leveraging Apple’s unified memory (up to 512 GB per Mac) and 1.2 TB/s memory bandwidth so the pair can hold the full model and stream token slices efficiently. Demonstrated agentic tasks include end‑to‑end iOS app changes and automated video review in DaVinci Resolve; these show on‑device capabilities for complex software engineering and frame‑by‑frame media analysis.
Weights and hosted access are available now across major clouds and platforms (Hugging Face, Amazon SageMaker AI, Google Cloud Model Garden, Microsoft Foundry Models) and through API partners (Baseten, Fireworks AI, Groq, Modal, Nebius, OpenRouter, Together AI). It runs on common inference engines (SGLang, vLLM, TensorRT‑LLM, HF Transformers, Apple MLX, llama.cpp, Unsloth) and ships with Helion GPU kernels that deliver 1.03-3.95× kernel speedups and a reported 3% end‑to‑end boost on AMD MI350. Safety post‑training met thresholds in Meta’s Superintelligence Scaling Framework; weights are released under the Muse Model License with limited commercial terms. Developers can further customize via PyTorch (TorchTitan) and the MLX adapter for Apple Silicon.
Summary generated by AI from the linked article. hn.today is not affiliated with Hacker News or Y Combinator.
Terence Tao discusses the allure of AI and large language models in advancing mathematics and science, comparing it to Borges's imagined Library of Babel containing all possible knowledge. He highlights how AI's potential to generate vast amounts of research could revolutionize scientific discovery, despite the challenges of sorting through misinformation. (terrytao.wordpress.com)
Companies rushed to adopt AI technology, leading to increased errors, low-quality work, and security risks. Some firms, like Meta, are now pulling back on AI integration to address these issues and reduce costs. (theguardian.com)
Accenture has limited employee spending on the AI tool Claude to $29 per month, down from as much as $1,000. Employees must now find ways to fund additional usage if they exceed the cap, reflecting broader industry cost-cutting measures on AI tools. (404media.co)
Flux 3 Image Klein is a 5.6 billion parameter model designed for text-to-image generation and image editing, accepting prompts and reference images to produce or modify images. It uses the Qwen3-VL-4B-Instruct text encoder and supports various applications, including fine-tuning and research. (huggingface.co)
Muse Spark Open is the most capable US open-weight multimodal model with 816 billion parameters, optimized for desktop use on Apple M5 Ultra Mac Studios. It excels at coding, complex reasoning, and agentic tasks, and is available through multiple platforms for customization and deployment. (research.meta.ai)
AI safety concerns are shared by both independent researchers and industry leaders, with many experts warning of potential existential risks. These warnings are not driven by corporate motives, as even prominent AI pioneers have expressed genuine concern about AI's dangers. (astralcodexten.com)
Today's best Hacker News stories, summarized and screenshotted, one email a day.