Meta just dropped a model that changes the conversation around open-weight AI. It is called Muse Spark Open, and at 816 billion parameters it is the largest open-weight model released by an American company this year. The kicker? It runs entirely on your desk, inside two Mac Studios, and it shipped iOS app fixes and edited video in DaVinci Resolve without touching a cloud server.
This is not the same Meta model we covered last week. Muse Wonder sits at the frontier and competes with the most powerful closed models from OpenAI and Anthropic. Muse Spark Open is a different bet entirely: an open-weight release optimised for developers who want serious AI capability without renting GPU clusters or sending data to third-party APIs.
What Makes Muse Spark Open Different
At 816 billion parameters, Spark Open is a multimodal model built for two specific workloads: coding and long-horizon agentic tasks. Those are precisely the areas where Western open-weight models have struggled to keep pace with Chinese alternatives, and Meta has been explicit that it designed Spark Open to close that gap.
On Artificial Analysis’ Intelligence Index, Spark Open scores 39. That puts it ahead of every other open-weight model from the West, but it still trails China’s GLM 5.3 at 45, Kimi K3 at 44, and Qwen3.8 at 40. The picture is similar on coding benchmarks: Spark Open matches or beats its Western open rivals on multimodal and knowledge tasks, but the Chinese models still write better code.
The Apple Partnership Changes the Deployment Story
Meta worked directly with Apple to tune Spark Open for the company’s MLX machine learning toolkit. The result is that two M5 Ultra Mac Studios with 1TB of combined memory can run the full model locally, something Meta describes as a first for a model in this size class.
Running a frontier-scale model on local hardware rather than cloud GPUs is significant for several reasons. Data never leaves your desk, so privacy-sensitive organisations can use the model without sending proprietary code or customer information to an API endpoint. The latency is effectively zero once the model is loaded, because there is no network round trip. The cost model also shifts from per-token API pricing to a fixed hardware investment.
Meta demonstrated this with real-world examples: the model shipped bug fixes to an iOS app and edited video footage inside DaVinci Resolve, all without cloud connectivity. For developers building AI-powered workflows into their products, that level of local capability has not been available at this scale before.
The Open-Weight Race Is Accelerating
Spark Open is the third major Western open-weight release this month alone. Reflection AI launched Beam earlier in October, and Mistral announced its 1-trillion-parameter Le Chonk, whose weights arrive on 27 October. The pace of releases has picked up dramatically since mid-2026, when it became clear that Chinese labs had opened a meaningful lead in accessible, high-performance open models.
Meta’s position is particularly interesting because the company is now playing both sides of the strategy. On one side, Muse Wonder competes at the absolute frontier behind a closed API, matching the most powerful models from OpenAI and Anthropic. On the other side, Muse Spark Open gives the open-weight community a genuine American contender in the space where China has dominated: large, capable models that anyone can download, fine-tune, and run on their own hardware.
The dual approach makes strategic sense. The closed frontier model captures the enterprise and high-end consumer markets where margin and control matter. The open model builds ecosystem, developer loyalty, and the kind of community-driven innovation that fed Meta’s earlier Llama success. Meta built its AI reputation on open Llama models before moving toward a more closed approach, and Spark Open represents a meaningful return to that original philosophy.
What This Means for Developers
If you are running an AI start-up or building AI features into a product, Spark Open changes the trade-offs you face. A model that fits on two high-end workstations and handles coding, agentic workflows, and multimodal tasks gives you a deployment option that was not available six months ago. You keep your data local, you control the model weights, and you are not paying per-token charges every time a user interacts with the system.
That said, the coding gap with Chinese models is real. If your primary use case is code generation, Kimi K3 or Qwen3.8 may still serve you better. Spark Open is strongest on multimodal reasoning and agentic planning, which is where Meta chose to focus its optimisation efforts.
For the broader AI industry, the significance is that the West has finally produced an open-weight model that can credibly compete in the 800B-parameter class. It does not lead the pack, but it is now in the race. With Apple hardware support enabling local deployment at a scale that was previously impossible, the practical barriers to running open frontier models have just come down.
Related Reading
- Meta’s Muse Wonder Crashes the Frontier Model Race: What It Changes
- Reflection AI’s Beam Is the West’s Latest Answer to China’s Open-Weight Dominance
- Why Western AI Labs Are Finally Getting Serious About Open-Weight Models
Meta built its AI reputation on open Llama models. With Spark Open, it is betting that the next generation of developers will remember that.
