Meta Platforms shares slipped on Thursday, October 15, as the company introduced Muse Spark Open, a new open-weight artificial intelligence model built for software development and automated tasks. Meta (META) stock closed the regular session at $736.72, down 0.11%, then fell a further 0.23% to $735.02 in after-hours trading. While the launch broadened developer access to advanced AI tools, the stock recorded losses in both sessions.

Meta Releases Muse Spark Open for Coding

Meta introduced Muse Spark Open as an open-weight model designed for software development and multi-step reasoning. The system carries 816 billion parameters, of which 42 billion are active during processing. That split matters for deployment: a model's full parameter set must be held in memory, while the active count shapes the computing power drawn during each processing step. It also supports context windows reaching one million tokens, enabling lengthy and complex coding assignments.

The model processes multiple information formats and handles programming tasks that require several rounds of decision-making. According to Meta, the system supports long-running projects and difficult software engineering work. Company evaluations also compared its performance against other open-weight models available to developers; because these results come from Meta's own testing, hands-on findings reported by developers through the release channels below may add independent reference points.

Open-weight releases allow developers to download model parameters and customize systems for specific operations. Businesses can therefore control deployment and manage information within their preferred environments. Meta released a smaller model, Muse Glimmer, in August before introducing this larger system.

Mac Studio Deployment Brings Hardware Demands

Meta designed Muse Spark Open to run locally on two Apple Mac Studios equipped with M5 Ultra chips. The setup relies on Apple's MLX framework to manage machine learning operations. However, this configuration requires substantial hardware resources, including approximately one terabyte of memory.

The company tested the model alongside Muse Code, its terminal assistant for development work. One exercise involved updating a Mastodon client through software fixes and new features. These demonstrations illustrated how the system manages connected tasks without repeated manual instructions.

Meta also tested the model with DaVinci Resolve, a professional video editing application. During that demonstration, the system reviewed footage and cropped selected sections of the video. The exercise showed the model's potential for desktop tasks outside conventional programming assignments.

Meta Expands Access Through APIs and Cloud Services

Meta made the model accessible through Baseten, Fireworks AI, Groq, Modal, Nebius, OpenRouter, and Together AI. Developers can download the weights from Hugging Face, Amazon SageMaker AI, Google Cloud, and Microsoft Foundry. These services provide alternatives to installing an expensive local hardware setup.

Compatible inference engines include vLLM, SGLang, TensorRT-LLM, llama.cpp, and Hugging Face Transformers. The new model also supports AMD, Apple, Intel, and Nvidia hardware for wider deployment. Developers can customize the system using PyTorch for specialized applications.

Open-weight distribution can reduce dependence on proprietary providers and increase deployment flexibility. Even so, running large models still demands technical expertise and significant computing resources. Developer uptake across these access routes may indicate how broadly the model sees practical use. Meta's release expands its AI offerings as Thursday's stock action remained slightly negative.

This article originally appeared on Blockonomi.