An 816B open-weight multimodal model for coding agents, long-context work and controlled deployment.

Muse Spark Open brings frontier-scale AI agents closer to local deployment

Date

Author

Infron Team

Category

Product Updates

Meta has released Muse Spark Open, the open-weight edition of its agentic and coding model. The release combines frontier-scale capacity, native multimodality, a one-million-token context window and a deployment story that reaches from hosted APIs and multi-GPU servers to Apple Silicon workstations.

The headline number is 816 billion parameters. The more important story is control. Developers and businesses can now evaluate a model in this capability class through an API, download its weights, customize it and decide where it should run.

Muse Spark Open is now available on Infron through an OpenAI-compatible API.

Muse Spark Open at a glance

Specification

Detail

Total parameters

816B

Active parameters

42B per token

Architecture

Sparse mixture of experts

Expert routing

256 experts per MoE layer, 8 selected per token

Context window

Up to 1,048,576 tokens

Inputs

Text, image, audio and video

Output

Text

Languages

Training data from more than 100 languages

Checkpoints

BF16 and NVFP4

Infron list price

$1.25 per 1M input tokens, $4.25 per 1M output tokens, $0.15 per 1M cached input tokens

The model uses a sparse mixture-of-experts architecture. Only a subset of the model is activated for each token, so per-token computation follows the 42B active parameter count rather than the full 816B total.

That efficiency does not remove the memory requirement. Serving the model still requires loading the full checkpoint, which is why hardware remains a central part of the Muse Spark Open story.

Built for agents rather than simple chat

Muse Spark Open is designed for coding and long-horizon agentic workflows. According to Meta, it can read a large codebase, plan changes, call tools, evaluate the result and correct itself when a step fails.

Its multimodal inputs allow the same model to work with source code, screenshots, documents, audio and video. This can reduce the need to connect a separate perception model to an agent before it can act on visual or audio information.

Meta reports testing fully local workflows that included shipping changes to an iOS application and operating DaVinci Resolve to review footage and produce vertical crops. These are vendor-reported examples rather than independent production benchmarks, but they make the intended use case clear: extended workflows with tools and state, not isolated question answering.

A one-million-token context window

Muse Spark Open supports up to 1,048,576 tokens of context. For agent systems, this can accommodate large code repositories, long documents, tool definitions and extended interaction histories.

Maximum advertised context is not the same as the context a team should use for every request. Longer prompts consume more input tokens, increase memory pressure and can affect latency. Production teams should test the model with representative workloads and use caching, retrieval and context management instead of automatically filling the entire window.

Infron lists cached input at $0.15 per million tokens, compared with $1.25 per million standard input tokens. Workflows that repeatedly send the same system instructions, tool definitions or repository context may benefit from prompt caching when the provider and request pattern qualify.

What local deployment actually requires

Meta describes Muse Spark Open as the first model in its class to run on a desktop using two Apple M5 Ultra Mac Studios with the MLX framework. This is technically meaningful, but the term desktop needs context.

The reference BF16 checkpoint contains about 1.65 TB of weights. Meta recommends budgeting roughly 2 TB of VRAM for BF16 serving once key-value cache, activations and runtime overhead are included. The NVFP4 checkpoint reduces the weight footprint to about 537 GB and can be served on a four-GPU NVIDIA Blackwell node.

Hugging Face also tested a three-bit quantization on an M3 Ultra with 512 GB of unified memory. The model occupied about 344 GB and generated approximately 21 tokens per second in the reported test using a 1,024-token prompt and 128 generated tokens.

Muse Spark Open therefore moves frontier-scale inference into workstation territory. It does not make that inference lightweight or inexpensive.

What it means for individual builders

Most developers will not begin by downloading hundreds of gigabytes of weights. Hosted inference is the practical entry point.

Individual builders can use Muse Spark Open to:

  • Test an agent against a real repository before choosing infrastructure.

  • Build multimodal tools that reason over screenshots, documents, audio or video.

  • Evaluate long-context behavior without operating a large GPU cluster.

  • Use an OpenAI-compatible client and keep application integration familiar.

The model is available on Infron under the identifier meta-ai/muse-spark-open. Current list pricing starts at $1.25 per million input tokens and $4.25 per million output tokens.

Ready to test it? Open Muse Spark Open on Infron and create an API key.

What it means for businesses

For businesses, the open-weight release expands the set of deployment and governance choices.

Teams can begin with hosted inference to establish quality, latency and cost on a representative workload. If the model proves valuable, they can then evaluate dedicated or self-hosted infrastructure for privacy, customization, predictable capacity or regional deployment requirements.

Potential business use cases include:

  • Software engineering agents working across large repositories.

  • Document agents that combine text, tables, screenshots and attachments.

  • Media operations that review and classify video or audio.

  • Internal assistants that need long context and tool access.

  • Domain-specific agents customized with proprietary data and workflows.

The ability to download weights does not automatically make self-hosting the most economical option. Hardware utilization, model serving expertise, observability, upgrades and incident response all contribute to total cost. A staged evaluation is usually more useful than choosing infrastructure from model specifications alone.

Evaluating production volume, private deployment or a custom commercial rate? Contact the Infron team.

Open-weight does not mean unrestricted

Muse Spark Open is released under the Muse Model License and Meta’s Acceptable Use Policy. Meta describes the license as offering limited commercial terms that allow many developers to use the model for free.

It should therefore be described as open-weight rather than fully open-source. Organizations planning redistribution, large-scale commercial deployment or model-based services should review the license and their intended use before adoption.

A broader change in the model market

The release narrows the gap between powerful closed APIs and smaller models that can be controlled locally.

Muse Spark Open does not eliminate that tradeoff. Its hardware requirements are substantial, and independent testing will be needed to establish how its launch claims translate to real workloads. It does, however, give developers another option: start with a hosted API, retain a path to the weights and move toward dedicated or local infrastructure when the economics and governance requirements support it.

The model decision is increasingly about more than benchmark rank. Teams must also decide where intelligence should run, which data should leave their environment and how much control they need over the serving stack.

Start building with Muse Spark Open

Muse Spark Open is live on Infron with an OpenAI-compatible API, published provider pricing and a path from initial testing to production deployment.

Sources

Useful links

Ready to explore Muse Spark Open further? These are the most useful places to start:

More Articles