Meta Muse Spark Open and open-weight AI: what businesses should know

Quick Go-Through:

Meta has released the weights of Muse Spark Open, an 816-billion-parameter model built for coding and agents. Here is what it can do, what the Muse Model License really allows, what running it costs, and how it compares with other open-weight models businesses can use today.

A big open-weight release from a US lab

On 15 October 2026 Meta released the weights of Muse Spark Open, a model it calls “the most capable US open-weight model to date.” Open-weight means anyone who accepts the license can download the trained model and run it on their own hardware or with a hosting provider of their choice, instead of only renting access through one company's API.

For businesses that is more than a headline. Open weights change who controls your data, how much you pay per task and how much you can customise the model. This article explains what Muse Spark Open is, what its license allows, what it takes to run, how it compares with other open-weight models released this month, and how to decide whether an open model belongs in your product.

What Meta actually released

According to Meta's announcement and model card, Muse Spark Open is a mixture-of-experts model with 816 billion total parameters, of which 42 billion are active for each token. It has a context window of one million tokens and is natively multimodal: it accepts text, images, video and audio as input and produces text as output. Meta says it is “built for coding and agentic work” and that it “excels at long-horizon tasks, complex software engineering, and multi-step reasoning.”

The model is derived from the hosted Muse Spark model Meta launched earlier this year, with extra safety training. The model card explains that the open version “has additional safety post-training to make it more robust against misuse as a standalone model without system guardrails,” and Meta recommends deploying it “as part of a broader AI system, with additional guardrails suited to the specific use case.”

Weights are available on Hugging Face, Amazon SageMaker AI, Google Cloud Model Garden and Microsoft Foundry Models, and hosted APIs are offered by Baseten, Fireworks AI, Groq, Modal, Nebius, OpenRouter and Together AI. It runs on common inference engines including vLLM, SGLang, TensorRT-LLM, Hugging Face Transformers, llama.cpp, Unsloth and Apple's MLX.

The Mac angle, and what it really costs to run

The most talked-about detail is that Meta worked with Apple so the model can run on two M5 Ultra Mac Studios using MLX. As The Verge reported, Mark Zuckerberg wrote on X that Muse Spark Open “runs great on new Mac Studios thanks to work with @johnternus and Apple,” and Apple CEO John Ternus replied: “Amazing to run a model of Muse Spark Open's caliber and scale this efficiently, right from your desk.”

Running on a desktop is not the same as running cheaply. The model card lists at least 700 GB of GPU memory for the 4-bit NVFP4 version and 2 TB for the full BF16 version. Meta's developer documentation puts the weight files alone at about 537 GB and 1.65 TB, with reference serving setups of four and eight Nvidia B300 GPUs. VentureBeat noted that Apple's 512 GB Mac Studio configuration had not shipped at launch, and that four of today's 256 GB models would cost $43,196 before tax.

For most small and mid-sized businesses the realistic route is a hosted API. VentureBeat reports that Together AI lists Muse Spark Open at $1.25 per million input tokens and $4.25 per million output tokens. That is cheaper than many frontier closed models, but several other open-weight models are cheaper still, so price alone does not make the case.

Read the license before you build

Meta describes the Muse Model License as having “limited commercial terms, which allow many developers to use the model for free.” The license text grants broad rights to use, modify, fine-tune, distill, distribute and sell the software, but with conditions that matter for a product team.

Attribution: you must prominently display “Built with Muse Spark Open” in the user interface of each product or service that uses the model or its derivatives, or in the accompanying documentation, website or settings menu.
Derivatives are defined broadly. They include models trained, fine-tuned or distilled using the software or its outputs, so the attribution and usage rules follow you if you train a smaller model on Muse Spark Open's answers.
Model-as-a-service threshold: if your company and its affiliates earn more than $20 million in gross revenue in any 12 months and you sell access to AI models through an API, you must notify Meta and sign a separate agreement first. Internal use, research and using the model to power features of your own product are exempt.
Usage policy: the license includes Meta's usage policy, which lists prohibited uses, and the software is provided as is, with Meta stating that it claims no ownership of outputs.

How it compares with other open-weight models

Muse Spark Open arrives in a crowded field. VentureBeat's Carl Franzen points out that Chinese labs such as Alibaba's Qwen, DeepSeek and Moonshot AI's Kimi have set the pace for open-weight AI, citing Hugging Face data that Chinese models accounted for 41% of downloads over the year to March 2026 and OpenRouter figures showing open-weight models handled 60% of token volume from US-originating requests in September.

On Meta's own benchmark table, Muse Spark Open leads its listed rivals on tool-use tests such as MCP-Atlas and JobBench, but trails Kimi K3 and GLM 5.3 on terminal coding and longer software-engineering tasks, and its reported Artificial Analysis Intelligence Index score of 39 sits below several Chinese models. Interconnects author Nathan Lambert welcomed the release anyway, writing on X: “I waited so long for US open weight models to be relevant again. Still more work to do!”

Licenses differ too. DeepSeek released DeepSeek-V4.1-Flash-1017 on Hugging Face this month, and its model card states that “this repository and the model weights are licensed under the MIT License,” with no attribution banner or revenue threshold. Meta's own smaller Muse Glimmer, a 30-billion-parameter model for local agents, uses Apache 2.0. For a business, the right comparison is not only which model scores highest, but which license, size and hosting options fit the job.

Why open-weight AI matters for businesses

1. Data control

You can run an open model on infrastructure you choose, including a private cloud region, so customer data does not have to pass through a model vendor's own API. VentureBeat also notes that where a model is developed does not by itself decide where your data is processed; the hosting choice does.

2. Customisation

Open weights can be fine-tuned on your own documents, tone and workflows. Meta points developers to PyTorch's TorchTitan for further tuning of Muse Spark Open.

3. Cost predictability

For steady, high-volume tasks, a self-hosted or competitively hosted open model can cost less per task than a premium closed model. For low or spiky volumes, a pay-as-you-go API is usually simpler.

4. No single-vendor lock-in

Because several providers serve the same weights, you can switch hosts if prices, limits or terms change, without rebuilding your product around a new model.

A practical checklist before choosing an open model

Start from the task: classification, search, support answers, coding help or document analysis need very different model sizes.
Test two or three candidates, open and closed, on your own real examples rather than relying on vendor benchmarks.
Read each license for attribution rules, revenue thresholds and usage restrictions, and keep a copy with your project files.
Decide on hosting early: a hosted API for speed, or your own servers when data control or volume justify it.
Add guardrails, logging and human review for anything customer-facing, as Meta itself recommends for Muse Spark Open.
Plan for change by keeping model calls behind one layer in your code so you can swap models as better ones arrive.

Where CoreCodify fits

Open-weight models like Muse Spark Open give businesses more choice, but choice only helps when the model is matched to a clear job and built into a product people enjoy using.

At CoreCodify we design and build websites, mobile apps, UI/UX and brands, and we build AI solutions such as assistants, smart search and document tools on top of both open and closed models. If you are weighing an open-weight model for your app or business, hire me for custom work at usamify.com.