← Friday 16 Octobernote 2 of 10
Open-weight AI modelsFri 16 Oct 2026 · 2 of 10

Meta.ai launches Muse Spark Open, a desktop‑capable multimodal AI model

Partly verified · 6/7 supported1 cited · 1 min read

meta.ai reports that Muse Spark Open is available today as the most capable US open‑weight model to date . meta.ai reports that the model is natively multimodal with 816 billion total parameters (42 billion active) and a 1 million‑token context window . meta.ai reports that it is built for coding and agentic work, excelling at long‑horizon tasks, complex software engineering, and multi‑step reasoning . meta.ai reports that Muse Spark Open is the first model in its class that can run on a desktop using two Apple M5 Ultra Mac Studios via the MLX framework, and it is available through API platforms including Baseten, Fireworks AI, Groq, Modal, Nebius, OpenRouter, and Together AI, with model weights downloadable from Amazon SageMaker AI, Google Cloud Model Garden, Hugging Face, and Microsoft Foundry Models . meta.ai reports that the model runs on inference engines such as SGLang, vLLM, TensorRT‑LLM, Hugging Face Transformers, Apple MLX, llama.cpp, and Unsloth, with optimizations for AMD, Apple, Intel, and NVIDIA silicon, and developers can further customize it using the PyTorch framework .

Why it matters. meta.ai reports that running a high‑performance multimodal model on desktop hardware expands access to advanced AI for home users and enables custom development with PyTorch .

Earlier on Open-weight AI models

Thu 15 Oct 2026WebKD: Minimal Editing of Failed Web Agent Episodes in Student’s StylearXiv reports that frontier web agents that rely on closed models incur high inference and computational costs. arXiv reports that smaller open models are a potential alternative but require training supervision from a human or teacher model. arXiv reports that the paper introduces WebKD, in which a teacher edits a student’s failed episode by locating the first mistake and rewriting the reasoning traces, states, and actions from that point onward. arXiv reports that WebKD aims to edit minimally, with the teacher writing each correction in the student’s own style, and that the teacher’s corrections can be more surprising to the student than the student’s own text.Thu 15 Oct 2026HiRO-ACEv2 delivers 3 km extreme‑weather simulations at 1/300th the cost of X‑SHiELDHiRO-ACEv2 is the newest update to the High Resolution Output AI2 Climate Emulator (HiRO-ACE) system. It simulates extreme rain, wind and pressure events at 3 km resolution with high accuracy, matching the target high‑resolution statistics of X‑SHiELD within 3% even for events that occur only once in a million data points. The system can generate global 3 km climate data at about one‑three‑hundredth the cost of the X‑SHiELD high‑resolution atmospheric model. HiRO‑ACE combines a coarse 100 km AI model (ACE) with a sharpening model (HiRO) that now uses the new downscaling component HiROv2, which predicts multiple high‑resolution fields simultaneously rather than only rainfall. HiROv2 also accurately estimates the total energy of tropical cyclones and can produce hurricanes with realistic eyes and rain.Tue 13 Oct 2026LFM2.5-24B-A2B on-device reasoning agent announced with expanded context and speedHugging Face announced LFM2.5-24B-A2B, a fast, private on-device agent model. The model scores 48.09 on LiveClaw, outperforming Qwen3.6-35B-A3B's 38.19 on the same benchmark. It uses a Mixture-of-Experts architecture with 40 layers, 64 experts per block, top‑4 routing and only 2 B active parameters per token, making reasoning tokens cheap. Inference runs at 184 tokens per second on an Apple M5 Max and up to 21 000 output tokens per second on a single NVIDIA B300 GPU. The context window expands to 131 K tokens and the vocabulary doubles to 128 K tokens, while the model generates an explicit chain of thought before answering; training scaled to ~33 T tokens with additional 2 T and 300 B token phases focused on reasoning, code, multilingual text and long‑context data, followed by five post‑training stages to become an agent.
Next · Open-weight AI modelsGranite Switch 4.1 unifies chat, decision‑making and speech on a VRAM‑efficient endpoint →