October 15: Meta Officially Launches Muse Spark Open— a native multimodal open-source model with 816 billion parameters,1M token contextand touted as the “strongest open weights in the US”. Unlike flagship cloud-only closed models, Muse Spark Open fully opens its weights and is specially optimized for Apple Silicon, enabling direct download and local execution on Mac. The sole trade-off: you’ll need at least Muse Wonder 1 TB of memory What Is Muse Spark Open?.
Muse Spark Open is an open-source model launched by Meta under its Muse series, co-optimized with Apple for macOS. Its positioning is crystal clear:
A heavy-duty open-source tool for programming and agent-based work. Meta’s official assessment calls it “the strongest open weights in the US to date” and publishes multiple benchmark results claiming superiority over current mainstream open-source models.— likely employs a Mixture-of-Experts (MoE) architecture; Meta has not disclosed the number of active parameters, but the total scale places it firmly in the top tier of open-source models.
core competencies
- 816 billion parameters— capable of ingesting ultra-long codebases or multiple documents at once, ideal for large-scale engineering tasks.
- 1-million-token context window— unified text-and-image modeling: writes code and understands screenshots, design mockups, and more.
- Native multimodalityDual Specialization in Programming and Agent Work
- — integrated with theMuse Code programming terminal to independently handle the full workflow—from bug fixing to complete project building.Local Execution on Mac
- Local Mac executionDeeply optimized for Apple Silicon; performs best on Mac Studio with M5 Ultra chip.
Meta also provided two real-world deployment cases: first, assigning the model to maintain code for a Mastodon client—including bug fixes, new feature implementation, and full builds; second, deploying it in professional video editing software DaVinci Resolve to automatically review and trim video footage. This demonstrates that Muse Spark Open is not just theoretical lab metrics but is already operational within actual workflows.
User experience and limitations
Open-source weights mean anyone can download, fine-tune, and commercialize them—no API fees—which is a major draw for both enterprises and individual developers. Yet its barrier to entry is equally demanding:
- Minimum 1 TB RAMAccording to Engadget, local execution requires at least 1 TB of RAM; standard PCs or consumer-grade MacBook models cannot run it—the most realistic option is the Mac Studio (M5 Ultra).
- Chinese-language capability undisclosedMeta’s official evaluation does not include a dedicated Chinese-language benchmark; actual user experience for Chinese speakers remains to be seen.
- High inference costEven with free weights, inference at the 816B scale demands far more compute power than typical self-hosted scenarios.
- Licensing details pending clarificationCommercial use eligibility and any restrictive clauses require careful review of the open-source license.
In short, Muse Spark Open is better suited for teams or developers with ample hardware budgets—not average individual users. For most people, its significance lies more in “raising the ceiling of open-source AI once again.”
Muse Spark Open vs. other open-source LLMs
| Model | Parameter count | Core Positioning | This Site’s Rating |
|---|---|---|---|
| Muse Spark Open | 816B | Strongest U.S. open-source model for coding and agentic tasks | 8.0 |
| Qwen3.8 2.4T A95B | 2.4T MoE | Alibaba’s flagship open-source model | 8.5 |
| DeepSeek V4 Pro | MoE | 1M context, flagship performance at an affordable price | 8.9 |
| Atria Dawn | 744B | Open-source agent | 8.7 |
| GLM 5.3 | — | Million-context reasoning model | 8.4 |
This table reveals: Muse Spark Open’s raw capabilities (816B parameters, multimodal, 1M context) are sufficient to place it in the top tier, but The 1TB memory requirementBrings its score back down to 8.0—DeepSeek V4 Pro and Atria Dawn It trades lower deployment costs for a higher overall score. If you prioritize “Made in the U.S. + maximum open-source capability” and have ample budget, it’s worth trying; otherwise, domestic open-source models remain the more cost-effective choice.
Overall Score
| 维度 | Score | evaluate |
|---|---|---|
| functional completeness | 9.0 / 10 | Native multimodality + 1M context + programming agent—comprehensive capabilities |
| 易用性 | 6.5 / 10 | 1TB memory requirement—cannot run locally on standard devices |
| Cost-effectiveness | 8.0 / 10 | Weights are free, but hardware investment is substantial |
| 中文支持 | 7.5 / 10 | Multilingual capability is solid; dedicated Chinese-language performance remains undisclosed |
| 输出质量 | 9.0 / 10 | Marketed as the strongest U.S.-based open-source model, with standout programming and agent performance |
Overall rating: 8.0/10
Frequently Asked Questions (FAQ)
What is Muse Spark Open?
Muse Spark Open is an open-source multimodal large language model released by Meta in October 2026, featuring 816 billion parameters and a 1-million-token context window, optimized for complex programming and agent tasks, with local execution optimizations for Apple Mac devices.
What configuration is required to run Muse Spark Open locally?
According to Engadget, local execution requires at least 1TB of memory, with the Mac Studio equipped with the M5 Ultra chip being the top recommendation. Standard PCs or consumer-grade MacBook models cannot handle it.
Is Muse Spark Open free?
The model weights are open-source and free to download, fine-tune, or even use commercially (subject to verification of the specific license terms). However, the hardware cost to run it is high—a machine with 1 TB of memory is expensive.
What is the difference between Muse Spark Open and Muse Wonder?
Muse Wonder is Meta’s closed-source frontier model, focused on active memory and Muse agents, accessible exclusively via cloud API; Muse Spark Open, by contrast, offers open-source weights optimized for local execution and programming capabilities—these two models serve entirely distinct purposes.
Does Muse Spark Open support Chinese?
Meta’s multilingual models generally deliver solid Chinese-language performance, but official evaluations do not separately disclose Chinese-language benchmark scores; Chinese users are advised to rely on their own testing for real-world experience.
Summarize
Muse Spark Open raises the ceiling for U.S.-based open-source models once again: 816B parameters, context windows up to 1 million tokens, native multimodality, and dual specialization in programming and agent intelligence—all concrete, tangible advancements. Yet its 1 TB memory requirement positions it more as a “team-grade” rather than a “personal-grade” tool. For well-funded developers, it represents an open-source option worthy of serious evaluation; for average users, it remains largely aspirational—domestically developed open-source models still offer better value for money.
Want to discover more great AI models? Check out our AI Model Library and Tool Comparison Engineor continue reading:Mistral Large 4 Evaluation · Reflection Beam Open-Source Model · All AI Evaluations.