Meta Muse Spark Open: Benchmarks, Features & Comparison

Meta's Muse Spark Open: 816 Billion Parameters, But Can It Compete with the AI Giants?

Meta's Muse Spark Open: A New Challenger in the AI Race, But How Does It Compare?

Published: 17 October 2026

Meta has released Muse Spark Open, an enormous new artificial intelligence model designed to compete with some of the most capable AI systems available today. With 816 billion parameters, advanced reasoning, multimodal capabilities and an impressive one-million-token context window, it certainly looks formidable on paper. But how does it actually perform against competitors such as OpenAI, Google, DeepSeek, Qwen and other leading AI models? And perhaps more importantly, what does its release mean for developers and AI enthusiasts?

The artificial intelligence industry certainly isn't slowing down. Just when we think we've reached another technological milestone, somebody comes along with a model that promises to push the boundaries a little further.

This time, it's Meta.

On 15 October 2026, Meta officially released Muse Spark Open, describing it as the most capable open-weight AI model developed in the United States to date.

Unlike the proprietary models powering services such as ChatGPT and Google's Gemini, Muse Spark Open makes its trained model weights available for developers to download, modify and deploy on their own infrastructure.

That is significant because access to the underlying model provides opportunities for experimentation, customisation and deployment that aren't always possible with proprietary AI services.

However, there's an important difference between open-weight and open-source, which we'll explore shortly.

And although Meta's announcement is impressive, the benchmarks reveal a rather more interesting story than the headline might suggest.

Source: Meta AI Research

What Exactly Is Muse Spark Open?

Muse Spark Open is a large language model developed by Meta Superintelligence Labs, the company's artificial intelligence research division.

It belongs to Meta's Muse family of AI models and is specifically designed for demanding applications involving reasoning, software development, autonomous agents and the interpretation of different types of information.

Its headline specifications are substantial:

Feature	Muse Spark Open
Developer	Meta Superintelligence Labs
Release date	15 October 2026
Total parameters	816 billion
Active parameters	Approximately 42 billion per token
Architecture	Mixture of Experts (MoE)
Context window	Up to 1 million tokens
Input capabilities	Text, images, audio and video
Output	Text
Reasoning levels	Low, medium, high and xhigh
Languages	Trained on data covering more than 100 languages
Model formats	BF16 and NVFP4
Availability	Downloadable weights and hosted APIs
Licence	Muse Model License

Sources: Meta's official announcement and Muse Spark Open model card.

Another interesting aspect is the model's architecture.

816 Billion Parameters, But Only 42 Billion Active?

At first glance, 816 billion parameters sounds almost absurdly large.

However, Muse Spark Open uses an architecture known as Mixture of Experts, or MoE.

Rather than activating every parameter whenever it processes information, the model uses a routing mechanism to select specialised neural networks, known as experts, for different tasks.

Muse Spark Open contains 256 experts within each MoE layer, but only eight are selected for each token.

Consequently, approximately 42 billion parameters are active during the processing of an individual token.

Think of it as employing 256 specialists but only calling upon the eight most relevant ones for a particular job.

This approach reduces computational requirements compared with activating all 816 billion parameters simultaneously.

There is, however, a catch.

Although only a fraction of the model is active at any given moment, the entire collection of parameters still needs to be stored and made accessible.

In other words, the model might be comparatively efficient when performing calculations, but it still requires an enormous amount of memory.

We'll come back to that rather inconvenient detail later.

What Can Muse Spark Open Actually Do?

Muse Spark Open is more than a conventional conversational chatbot.

Meta has specifically trained it for what the industry calls agentic AI, where a model can work through complicated tasks involving planning, external tools and multiple stages of execution.

For example, rather than simply answering a programming question, an AI agent could inspect a software project, identify a problem, modify the relevant files, run tests and attempt to correct any errors.

Muse Spark Open is designed to support these workflows.

Its capabilities also extend beyond conventional text processing.

The model can interpret images, analyse audio recordings, examine video footage and combine these different inputs when reasoning about a task.

Imagine providing an AI assistant with a screen recording of a software problem and asking it to investigate the issue, explain the likely cause and suggest a solution.

That's the sort of interaction Muse Spark Open is designed to handle.

Its one-million-token context window is another notable feature.

A token is a small unit of text or other encoded information processed by an AI model. A context window this large can accommodate extensive documentation, substantial codebases or lengthy conversations, although images, audio and video also consume context capacity.

Of course, being able to accept a million tokens doesn't automatically mean the model will reason perfectly over every piece of information contained within them.

Context capacity and the ability to use that context accurately are two different things.

How Does Muse Spark Open Compare With Other AI Models?

This is where things become quite interesting.

Meta has released benchmark results comparing Muse Spark Open against several competing open-weight models, including Kimi K3, GLM 5.3 and Alibaba's Qwen3.8.

However, we also have independent evaluations from Artificial Analysis, allowing us to compare it against proprietary models from OpenAI, Anthropic and Google.

And the results demonstrate why we should never judge an AI model by its parameter count alone.

Overall Intelligence: The Benchmark Results

Artificial Analysis is an independent AI evaluation platform that assesses models across reasoning, coding, knowledge, agentic workflows and other capabilities.

Its Intelligence Index combines multiple benchmarks to provide a broader measurement of model performance.

Here are selected results from its Intelligence Index, version 4.3.2, as available on 16 October 2026.

AI Model	Intelligence Index
Claude Opus 5.5 (Max)	58
GPT-6 Astra (Max)	53
Gemini 4 Argon (High)	53
Meta Muse Spark 1.3 (xhigh)	45
GLM 5.3 (max)	45
Kimi K3 (max)	44
Qwen3.8 2.4T-A95B	40
Muse Spark Open (xhigh)	39
Mistral Large 4 Preview	38
DeepSeek V4 Pro 0813 (Max)	36

Important: These are composite index scores, not percentages. The table includes models with different reasoning settings and access arrangements. They should be treated as a broad comparison rather than a definitive measure of performance on every task.

Source: Artificial Analysis, including comparisons with GPT-6 Astra, Claude Opus 5.5 and Gemini 4 Argon.

What Do These Numbers Tell Us?

Firstly, Muse Spark Open is clearly a competitive model.

Its score of 39 places it close to several established open-weight competitors, including Alibaba's Qwen3.8 and Moonshot AI's Kimi K3.

It also scores above the evaluated DeepSeek V4 Pro configuration.

However, it remains behind the most capable proprietary models in this particular index.

Claude Opus 5.5, GPT-6 Astra and Gemini 4 Argon all achieve substantially higher composite scores.

Perhaps more surprisingly, Meta's proprietary Muse Spark 1.3 also scores higher than Muse Spark Open, demonstrating that opening a model's weights does not automatically mean releasing the company's most capable model.

But that doesn't diminish what Meta has achieved.

A downloadable model capable of approaching the performance of major proprietary systems gives developers something quite valuable. Much greater control over how their AI operates!

And when we examine individual benchmarks, Muse Spark Open demonstrates some interesting strengths.

Agentic AI: Where Muse Spark Open Shows Promise

One of Muse Spark Open's main selling points is its ability to operate as an autonomous AI agent.

Rather than simply generating text, an agent can interact with external systems, execute commands, retrieve information and work towards completing a task.

Meta's published evaluations provide several interesting comparisons.

Benchmark	Muse Spark Open	GLM 5.3	Kimi K3
JobBench	61.8	61.4	54.3
MCP-Atlas	89.2	84.2	82.3
Toolathlon Verified	72.8	73.0	76.5
OSWorld Verified	80.4	—	84.8

Higher scores indicate better performance within each benchmark. Missing results are indicated by a dash.

Source: Muse Spark Open official model card. Results combine independent, provider-reported and Meta-run evaluations.

Why Is MCP-Atlas Interesting?

MCP stands for Model Context Protocol, a standard that allows AI models and applications to communicate with external tools and data sources.

For example, an AI assistant might use MCP-compatible tools to search files, access databases, retrieve information or interact with development environments.

Muse Spark Open scores 89.2 on MCP-Atlas, compared with 84.2 for GLM 5.3 and 82.3 for Kimi K3.

That is a meaningful result for developers building applications around external tools.

It suggests that Muse Spark Open is capable in the tested tool-interaction scenarios.

However, real-world reliability will still depend on the quality of the tools, the agent framework and how successfully the system handles errors.

An impressive benchmark score is encouraging, but nobody wants an autonomous coding agent cheerfully deleting the wrong directory because it misunderstood an instruction.

Some things are best left with a confirmation button!

Coding Performance: Not Quite the Victory Meta Might Have Hoped For

Given Meta's emphasis on software engineering, coding performance deserves particular attention.

Two relevant benchmarks are DeepSWE and Terminal-Bench.

DeepSWE evaluates an AI agent's ability to complete complex software engineering tasks.

Terminal-Bench measures performance on technical tasks conducted through a command-line environment.

Here are Meta's published comparisons:

Coding Benchmark	Muse Spark Open	GLM 5.3	Kimi K3
DeepSWE 1.1	60.0	67.0	68.0
SWE-Atlas-QnA	58.0	59.0	66.0
Terminal-Bench 2.1	76.4	83.9	85.0

Source: Meta's published model evaluations.

These results are revealing.

Despite being marketed as a model optimised for coding and agentic workflows, Muse Spark Open trails both GLM 5.3 and Kimi K3 on these particular software engineering benchmarks.

That doesn't mean it is a poor coding model.

It simply means there are competing models demonstrating stronger results under these evaluation conditions.

There is another important complication.

Coding benchmarks frequently depend on the surrounding agent framework, sometimes called a harness, which determines how the model receives instructions, accesses tools and executes tasks.

Meta's evaluation methodology acknowledges that some competing models were tested using different coding harnesses.

Consequently, these results don't isolate the underlying model's capabilities perfectly.

Nevertheless, the differences are substantial enough to warrant attention.

An Interesting Independent Result

Artificial Analysis also reports results for the newer Terminal-Bench 4.0 evaluation.

Here, Muse Spark Open achieves just 8%, compared with 59% for GPT-6 Astra (Max) and 60% for Claude Opus 5.5 (Max).

That is a much larger difference.

However, Terminal-Bench 4.0 is a different benchmark version from Terminal-Bench 2.1, so the 8% and 76.4 results should not be directly compared.

The wider point remains: performance can vary dramatically depending on the task, benchmark version and agent environment.

Anyone selecting an AI model for serious software engineering should test it against their own workflows rather than relying entirely on promotional benchmark tables.

Source: Artificial Analysis comparison.

Reasoning and Knowledge: A More Encouraging Picture

Muse Spark Open performs very well on several reasoning and knowledge evaluations.

Meta reports the following results:

Benchmark	Muse Spark Open	GLM 5.3	Kimi K3
HLE-Diamond (without tools)	26.1	16.4	22.2
AA-Omniscience	25	14	20
SciCode	58.5	59.0	59.5
GPQA Diamond	89.0	91.7	93.5

Source: Meta's benchmark results.

HLE-Diamond is a challenging subset of Humanity's Last Exam, designed to test advanced knowledge and reasoning.

AA-Omniscience is interesting because it rewards correct answers while penalising incorrect ones.

In other words, confidently inventing an answer is not a good strategy.

Muse Spark Open's score of 25 compares favourably with the competing open-weight models shown here.

For anyone concerned about AI hallucinations, that's encouraging.

However, a strong result on a knowledge-reliability benchmark should not be interpreted as a guarantee that the model will consistently provide accurate answers.

Like every large language model, it can still make mistakes.

Multimodal Capabilities: More Than Just Text

Another important feature of Muse Spark Open is its native multimodal architecture.

The model can process text, images, audio and video within the same system.

This provides considerable flexibility for applications involving visual interpretation, speech analysis, document understanding and automated workflows.

For example, a software development assistant could potentially inspect both source code and screenshots of an application's behaviour.

Meta's multimodal evaluations include:

Benchmark	Muse Spark Open	Kimi K3
ZeroBench (with tools)	50.0	41.0
BabyVision (with tools)	80.2	85.7
SimpleVQA (with tools)	73.6	71.9
IntPhys2	68.3	60.3
VideoMMEv2	61.5	65.4

Source: Meta's published evaluations.

These results show a mixed but competitive picture.

Muse Spark Open performs well on ZeroBench, SimpleVQA and IntPhys2, while Kimi K3 scores higher on BabyVision and VideoMMEv2.

One important limitation is worth mentioning.

Although Muse Spark Open can understand audio and video inputs, its output is text.

It is not a direct replacement for dedicated speech synthesis, audio generation or video generation models.

However, its ability to interpret those media types could make it a useful component within larger creative or automated production pipelines.

For example, an AI system might use Muse Spark Open to analyse video footage, identify interesting moments and make editing decisions, while other specialist tools perform the actual rendering.

That's where multimodal reasoning becomes quite useful.

Running Muse Spark Open Locally: You'll Need More Than a Gaming PC

One of the most interesting aspects of Meta's announcement involves an unexpected collaboration with Apple.

Meta has worked with Apple to optimise Muse Spark Open for the M5 Ultra Mac Studio, using Apple's MLX machine-learning framework.

According to Meta, the complete model can run locally across two Mac Studio machines equipped with M5 Ultra processors.

Each machine can support up to 512GB of unified memory, providing a combined capacity of approximately 1TB.

Apple's unified memory architecture allows its processors to access large quantities of memory without relying on conventional discrete GPU configurations.

This makes such systems interesting for running exceptionally large AI models.

And Muse Spark Open certainly qualifies as exceptionally large.

Source: Meta AI Research.

What About NVIDIA GPUs?

Meta also provides deployment options for NVIDIA and AMD hardware.

According to its model card, conventional GPU deployments require approximately:

Model Format	Published GPU Memory Requirement
NVFP4	At least 700GB
BF16	At least 2TB

The model is also available through frameworks including vLLM, SGLang, TensorRT-LLM and llama.cpp.

These requirements make it impractical to run the complete model on a typical enthusiast gaming PC.

Even powerful consumer GPUs with 24GB or 32GB of VRAM fall substantially short.

More aggressive quantisation and system-memory offloading could potentially change deployment possibilities, but that would involve additional engineering and potentially significant performance compromises.

For most individual developers, accessing Muse Spark Open through a hosted API will be considerably more practical than attempting to deploy the complete model locally.

Source: Muse Spark Open model card.

Is Muse Spark Open Actually Open Source?

This is something that deserves more attention than it receives.

Meta describes Muse Spark Open as an open-weight model.

That means developers can access its trained parameters and use them within the conditions of its licence.

They can also fine-tune the model, create derivatives and deploy it in their own applications.

However, providing model weights is not necessarily the same as releasing an entirely open-source AI system.

A fully reproducible model would require considerably more information, potentially including training data, training procedures and other components used during development.

Muse Spark Open is distributed under Meta's own Muse Model License rather than a conventional permissive software licence such as MIT or Apache 2.0.

What Are the Licensing Restrictions?

Meta's licence permits a wide range of uses, including commercial applications.

However, it contains several conditions.

One notable restriction concerns companies offering AI models to third parties as a service.

If a company and its affiliates generate more than $20 million in gross revenue during any consecutive twelve-month period, they must notify Meta and enter into a separate agreement before offering Muse Spark Open as part of a Model-as-a-Service business.

Importantly, this restriction does not generally apply to ordinary internal commercial use or to companies incorporating the model into features of their own products that are not Model-as-a-Service offerings.

The licence also requires appropriate attribution, including displaying Built with Muse Spark Open in the product interface or accompanying materials.

For independent developers and smaller businesses, the licence provides considerable flexibility.

For companies offering commercial AI hosting services, the additional conditions may become more significant.

Therefore, Muse Spark Open is an important open-weight release, but it should not be confused with an unrestricted open-source model.

Source: Official Muse Model License.

Performance Versus Cost: Another Important Consideration

Model intelligence is only one part of the equation.

For developers building AI applications, inference costs and response speeds can be equally important.

Artificial Analysis reports the following hosted API pricing for Muse Spark Open:

Pricing Metric	Muse Spark Open
Input tokens	$1.25 per million
Output tokens	$4.25 per million
Cached input tokens	$0.15 per million
Measured output speed	Approximately 102 tokens per second

These are observed hosted-service figures, not the costs of operating a local deployment. Actual pricing and performance vary by provider, configuration and workload.

For comparison, Artificial Analysis lists GPT-6 Astra at $10 per million input tokens and $50 per million output tokens.

That is a substantial pricing difference.

However, GPT-6 Astra also achieves considerably higher overall benchmark results.

Choosing between them therefore depends on the workload.

For applications where the additional reasoning performance matters, the higher cost might be justified.

For applications involving large quantities of relatively straightforward tasks, the lower inference costs of Muse Spark Open could make it attractive.

The most expensive model isn't automatically the most economical, and the cheapest model isn't necessarily the best value.

A model that requires repeated attempts to complete a task successfully can ultimately cost more than one that gets it right the first time.

Source: Artificial Analysis.

Why Does Muse Spark Open Matter?

It's tempting to concentrate entirely on benchmark rankings when discussing new AI models.

After all, the industry has become rather obsessed with comparing scores.

But the significance of Muse Spark Open extends beyond whether it can outperform a particular version of Claude, Gemini or GPT.

There are several broader developments worth considering.

Firstly, it demonstrates that extremely capable AI systems can be made available outside proprietary cloud platforms.

Secondly, its mixture-of-experts architecture illustrates how developers are attempting to balance enormous model capacity against practical computational requirements.

Thirdly, the collaboration between Meta and Apple highlights the growing importance of hardware and software optimisation.

And finally, it increases the range of options available to researchers and developers who want more control over their AI systems.

That last point is pretty important.

When organisations depend entirely on proprietary AI services, they also depend on the provider's pricing, availability, usage policies and technical decisions.

Open-weight models offer an alternative, even if running the largest examples still requires substantial infrastructure.

Of course, open weights introduce their own challenges.

Developers become responsible for deployment security, resource management, safety controls and ongoing maintenance.

Meta has also applied additional safety training to this release and recommends that developers implement appropriate safeguards around tools and autonomous actions.

Access to powerful technology is valuable, but deploying it responsibly remains essential.

My Thoughts: An Important Release, But Not a Revolution

Muse Spark Open is undoubtedly an interesting addition to the rapidly expanding AI landscape.

Its architecture is sophisticated, its multimodal capabilities are impressive, and its performance across several agentic and reasoning benchmarks demonstrates substantial engineering progress.

However, the available evidence does not support describing it as the most capable AI model overall.

Leading proprietary systems continue to outperform it on several important evaluations.

Even within the open-weight ecosystem, competitors such as GLM 5.3, Kimi K3 and Qwen3.8 demonstrate higher overall intelligence scores.

And despite Meta's emphasis on coding, some competing models achieve stronger results on complex software engineering benchmarks.

Nevertheless, I think focusing exclusively on which model occupies the top position misses something important.

AI development isn't just about building increasingly intelligent models. It's also about making that intelligence more accessible, adaptable and practical.

Muse Spark Open makes progress towards that goal, although its enormous memory requirements mean truly local deployment remains beyond the reach of most enthusiasts.

The model also demonstrates something we've seen repeatedly throughout AI development and that's parameter count alone tells us remarkably little about real-world performance.

An 816-billion-parameter model isn't automatically better than a smaller one.

Architecture, training quality, reasoning techniques, tool integration and inference optimisation all contribute to the final result.

And as the benchmarks demonstrate, a model can perform exceptionally well in one area while falling behind in another.

For developers, that means choosing the right model for the right task rather than automatically selecting whichever one has the most impressive specifications.

For AI enthusiasts, it means we're gaining more opportunities to experiment with technology that would have seemed extraordinary only a few years ago.

And for the wider industry, Muse Spark Open provides yet another reminder that competition in artificial intelligence remains intense.

Final Thoughts

Perhaps the most encouraging aspect of Muse Spark Open isn't its enormous parameter count or even its benchmark performance.

It's the continuing expansion of the open-weight AI ecosystem.

We're reaching a point where sophisticated AI capabilities are no longer exclusively available through a handful of proprietary services.

There are still substantial technical and financial barriers, particularly when models require hundreds of gigabytes of memory.

But progress in quantisation, inference optimisation and specialised hardware continues to make increasingly capable models more accessible.

Muse Spark Open isn't the final word in artificial intelligence.

Given the speed at which this industry is moving, it probably won't even be the final word this month.

But it represents another significant development in the evolution of AI.

And as someone who enjoys experimenting with models, pushing hardware beyond its intended limits, I find that rather exciting.

The interesting question isn't simply how intelligent AI models will become. It's how much of that intelligence we'll eventually be able to run, customise and control ourselves.

And that is a development worth watching.