Join our Newsletter — 33% off our NHI Course
Home› Guides› Meta Muse Spark Open: Specs, Context Window, Pricing,…
Guide AI Platforms

Meta Muse Spark Open: Specs, Context Window, Pricing, Benchmarks and Availability

PDF for newsletter subscribers · free to read online

← All guides
By NHI Mgmt Group, edited by Lalit Choda Updated 15 November 2026 14 min read
On this page

Muse Spark Open is the open-weight release of Muse Spark from Meta Superintelligence Labs, an 816B-parameter mixture-of-experts model with 42B active parameters, released on 15 October 2026 under the Muse Model License. This profile collects the published facts a team needs before it downloads, hosts or buys access to the model: architecture, context window, weight formats and hardware, licence conditions, gated access, hosted providers, reasoning settings, provider-reported benchmarks and the safety results in the model card. It is written for developers, platform engineers, buyers and AI governance teams. Every fact comes from Meta's own announcement, developer documentation and Hugging Face model card, listed under Standards and references. Where Meta publishes nothing on a point, the page says so.

Last checked: 2026-11-15.

This is an independent reference page from NHI Mgmt Group based on public provider documentation; it is not sponsored or endorsed by any provider. It uses Meta's public Muse Spark Open documentation as of November 2026. Model details change, so confirm against Meta's current documentation.

Key takeaways

  • 816B total, 42B active parameters. A sparse mixture-of-experts transformer with 256 experts, 8 routed per token, and a built-in 1.8B-parameter perception encoder.
  • 1,048,576-token context window. Input is text, image, video and audio; output is text. Meta does not state a maximum output or knowledge cutoff on the pages checked.
  • No Meta Model API ID and no Meta price. The weights are downloaded and self-hosted, or used through hosted providers such as Baseten, Fireworks AI, Groq, Modal, Nebius, OpenRouter and Together AI, each with its own pricing.
  • Muse Model License with a revenue condition. Model-as-a-Service providers above US $20 million in gross revenue over any 12 months must notify Meta and sign a separate agreement, and products must display "Built with Muse Spark Open".
  • Gated download. The Hugging Face form asks for full legal name, date of birth and organisation name, cannot be edited after submission, and the model is not intended for users under 18.

What is Muse Spark Open?

Muse Spark Open is the open-weight member of Meta's Muse family. Meta's research post of 15 October 2026 calls it "the open-weight version of Muse Spark" and describes it as natively multimodal and focused on coding and agentic work, including long-horizon tasks, software engineering and multi-step reasoning. Meta's developer documentation calls Muse Spark "Meta's agentic and coding model" and lists the open release for self-hosted or provider-hosted agentic and coding work and for post-training with supervised fine-tuning or reinforcement learning.

Unlike the closed Muse Spark, Muse Spark Open is not served on the Meta Model API: Meta's models page shows no model ID for it. Meta does not state which Muse Spark version (1.1, 1.2 or 1.3) the open weights correspond to. The research post also says the model runs on two Apple M5 Ultra Mac Studios using MLX, which it describes as "the first model in its class to run on your desktop". Meta's other open-weight Muse model is the smaller Muse Glimmer, distilled from Muse Spark. For the hosted model, see the Meta Muse Spark profile; for the full line-up, see Meta Models.

Key facts at a glance

As of 2026-11-15. Source: Meta research post, Muse Spark Open documentation, models page and Hugging Face model card.

FieldValueSource
ProviderMeta (Meta Superintelligence Labs)Hugging Face model card
Repository IDsmeta-models/Muse-Spark-Open-816B-A42B (BF16); meta-models/Muse-Spark-Open-816B-A42B-NVFP4Hugging Face model card
Meta Model API IDNone (not served on the Model API)Models page
Release date2026-10-15Research post; model card
StatusCurrentModels page
Context window1,048,576 tokensMuse Spark Open documentation
Max outputNot stated by Meta on the pages checked (2026-11-15)Model card
Knowledge cutoffNot stated by Meta on the pages checked (2026-11-15)Model card
Input modalitiesText, image, video, audioModels page; model card
Output modalitiesTextModels page; model card
Open weightsYes, gated (BF16 and NVFP4)Model card; documentation
LicenceMuse Model License (Muse Spark Open version, dated 15 October 2026) plus Usage PolicyModel card
Parameters816B total, 42B active (the Hugging Face page header shows 822B)Model card; documentation
LanguagesTrained on data from more than 100 languagesDocumentation; model card

Architecture

As of 2026-11-15. Source: Hugging Face model card.

ComponentValue
TypeMixture-of-experts transformer with Perception Encoder
Layers and hidden dimension62 layers, 8,192 hidden dimension
Experts256, with 8 activated per token (top-k routing)
AttentionGrouped-query attention, 128 query heads, 16 KV heads, head dimension 64; gated attention with QK-norm
Vocabulary202,048 tokens; input and output embeddings untied
Perception Encoder1.8B parameters, 50 layers, 1,536 hidden dimension, 4,096 output dimension

Release and version history

As of 2026-11-15. Source: Meta research post and Hugging Face model card.

DateEventSource
2026-10-15Muse Spark Open released with weights on Hugging Face and launch partners for hosting, tooling and siliconResearch post; model card
2026-10-15Muse Model License (Muse Spark Open version) datedModel card
Not statedFull evaluation report, described as "available soon"Research post

There is one published checkpoint, in two precisions. Meta lists no point releases or revised weights on the pages checked (2026-11-15). The closed Muse Spark it derives from has versions 1.1, 1.2 and 1.3 on the Model API.

How much does Muse Spark Open cost?

As of 2026-11-15. Meta does not sell Muse Spark Open by the token. Cost is either the hardware to run it or a hosted provider's price.

ItemPriceNotesSource
Weights downloadNo fee stated by MetaMeta describes "limited commercial terms, which allow many developers to use the model for free"Research post
Meta Model APINot offeredNo Model API IDModels page
Hosted providersSet by each providerBaseten, Fireworks AI, Groq, Modal, Nebius, OpenRouter, Together AI; Meta does not publish their pricesDocumentation; research post
Large Model-as-a-Service providersSeparate written agreement with MetaRequired above US $20 million aggregate gross revenue in any consecutive 12 monthsModel card (licence)
Self-hostingOperator's hardware and operationsSee hardware figures under Context window and output limitsDocumentation

The licence excludes internal use, research and features inside a company's own product from the revenue condition, per the model card. Read the licence text itself before relying on that summary. See the LLM Pricing Tracker.

Context window and output limits

Meta's documentation states a 1,048,576-token context window, and the model card states 1M tokens at inference. Self-hosting at that length is a hardware question: the memory figures below include KV cache only where Meta says so.

LimitValueNotesSource
Context window1,048,576 tokens1M tokens at inferenceDocumentation; model card
Max outputNot stated by Meta on the pages checked (2026-11-15)NoneModel card
BF16 weightsAbout 1.65 TB; roughly 2 TB VRAM to serve2 TB includes KV cache, activations and runtime overhead; eight GPUs such as 8 x B300Documentation
NVFP4 weightsAbout 537 GB; at least 700 GB VRAMMostly 4-bit with some BF16 layers; a single 4-GPU NVIDIA Blackwell nodeDocumentation; model card
Desktop optionTwo Apple M5 Ultra Mac Studios with MLXEach with up to 512 GB unified memoryResearch post
Video inputAt least 448 px on the short side recommendedShort or low-resolution videos (about 100 px) may be rejectedModel card

See Mixture of Experts Explained and Context Windows Explained.

Modalities and tools

  • Inputs: text, image, video and audio, handled by the built-in perception encoder; output is text.
  • Reasoning effort: set through reasoning_effort or chat_template_kwargs with values low, medium, high or xhigh; xhigh is the default. Meta's deployment page states reasoning cannot be disabled (requests with "none" are rejected), and max_tokens must cover both reasoning and the final answer.
  • Recommended sampling: temperature 1.0, top_p 1.0, top_k disabled.
  • Tool use: the model card describes sequential tool invocation across multi-turn workflows, including recovery from failed tool calls. Meta's prompting guide covers chat, system prompts, reasoning, tool calling and media input.
  • Agentic tests: the research post describes on-device agentic workflows, including iOS changes to the Ice Cubes app and driving DaVinci Resolve.
  • Inference engines: SGLang, vLLM, TensorRT-LLM, Hugging Face Transformers, Apple MLX, llama.cpp and Unsloth, with optimisations for AMD, Apple, Intel and NVIDIA silicon.
  • Post-training: PyTorch, including TorchTitan; Meta's documentation has a post-training section for SFT and RL.

An open-weight agent model that calls tools carries the credentials of every tool it can reach. Scope each integration, log tool calls and follow Meta's own advice to require human confirmation for irreversible actions. See the Open-Weight Model Security Guide.

Where is Muse Spark Open available?

As of 2026-11-15. Source: Muse Spark Open documentation and research post. Providers listed alphabetically within each row.

ChannelDetailsSource
Weight downloadsMeta and Hugging Face (gated)Documentation
Model cataloguesAmazon SageMaker AI, Google Cloud Model Garden, Microsoft Foundry Model CatalogDocumentation
Hosted inference platformsBaseten, Fireworks AI, Groq, Modal, Nebius, Together AIDocumentation
GatewayOpenRouterDocumentation
Meta Model API, meta.ai, Meta AI appNot offered for Muse Spark Open on the pages checked (2026-11-15)Models page
RegionsNot stated by Meta on the pages checked (2026-11-15)Documentation

Meta names these as launch partners. Each provider sets its own terms, data handling and regions. See Self-Hosted LLMs vs Hosted LLM APIs.

Access restrictions

Muse Spark Open is an open-weight model with gated access. To download from Hugging Face, a user must log in and complete a form with full legal name, date of birth and full organisation name, which cannot be edited after submission. The model card states the model is not intended for users under 18. Use is governed by the Muse Model License and its Exhibit A Usage Policy, which prohibits uses including illegal activity, harassment, unlicensed professional practice, malware, weapons, self-harm, fraud, disinformation and impersonation. The licence requires products to prominently display "Built with Muse Spark Open" in the UI, documentation, website or settings, and sets the separate-agreement condition for large Model-as-a-Service providers described above.

Benchmarks (provider-reported, dated)

As of 2026-11-15. All figures are provider-reported by Meta in the Hugging Face model card (October 2026), at xhigh reasoning effort. The Artificial Analysis Intelligence Index figure is as reported by Meta in the card and was not checked on Artificial Analysis's own pages for this profile.

EvaluationMuse Spark OpenSource and date
Artificial Analysis Intelligence Index39Model card, 2026-10
MCP-Atlas89.2Model card, 2026-10
OSWorld Verified80.4Model card, 2026-10
Terminal-Bench 2.176.4Model card, 2026-10
Toolathlon Verified72.8Model card, 2026-10
DeepSWE 1.160Model card, 2026-10
SWE-Atlas-QnA58Model card, 2026-10
GPQA Diamond89.0Model card, 2026-10
HLE-Diamond (no tools)26.1Model card, 2026-10
Humanity's Last Exam (with tools)61.0Model card, 2026-10
ZeroBench (with tools)50.0Model card, 2026-10
VideoMMEv261.5Model card, 2026-10

The card sets these against other open-weight models (GLM 5.3, Kimi K3, Mistral Large 4, Nemotron 3 Ultra, Qwen3.8 and others); those comparator scores are not reproduced here. The research post's "Key Benchmarks" section gives no figures in its text. See How to Read LLM Benchmarks.

Model card and system card

Meta publishes a model card on Hugging Face and refers to an evaluation methodology report; a full evaluation report is "available soon" per the research post. No separate system card is linked on the pages checked (2026-11-15). Meta states the model received "additional safety post-training" because of open-weight considerations and was evaluated under the Meta Superintelligence Scaling Framework, meeting its release thresholds. What the card states:

  • Cybersecurity: refused 99.2% of jailbreaks on Meta's internal benchmark and 94.8% across aggregated public benchmarks, without system-level safeguards.
  • Chemical and biological: refused over 98% of harmful requests on SecureBio's BioTIER and over 99% on an internal single-turn chemistry benchmark.
  • Misalignment: 98.1% honesty on MASK and a 0% misbehaviour rate on Agentic Misalignment.
  • Deployment advice: deploy within a broader system with additional guardrails, and require human confirmation for irreversible agentic actions.
  • Limitations: may produce inaccurate, biased or objectionable responses and make multi-step reasoning errors, especially in novel scenarios; performance may degrade outside strongly supported languages; developers are responsible for their own safety testing.

See Model Cards and System Cards Explained and the Frontier AI Model Risk Frameworks Guide.

Data handling

  • Self-hosted weights: prompts and outputs stay in the operator's environment; Meta states no inference data flow for downloaded weights on the pages checked.
  • Hosted providers: retention, training use and residency are set by each provider, not by Meta.
  • Download form: the gated Hugging Face form collects legal name, date of birth and organisation; the Meta download route requires sign-in and licence acceptance, after which the model runs offline with "no API key", per Meta's get the model page.
  • Training data: not described by Meta on the pages checked (2026-11-15), beyond more than 100 languages.

Self-hosting notes from Meta's deployment guide

  • Network exposure: several example server configurations bind to 0.0.0.0; TensorRT-LLM defaults to localhost.
  • Authentication: the llama.cpp API key is optional and its /health and /v1/models endpoints answer without authentication; the Unsloth API requires a bearer key.
  • Remote code: the recommended TensorRT-LLM configuration sets trust_remote_code to true.
  • Integrity: the BF16 download has 65 weight shards and NVFP4 127; Meta warns a runtime cannot load the checkpoint if any file is missing or renamed, and the page describes no checksums, so record your own hashes.
  • Served context: TensorRT-LLM's validated serving window is 49,152 tokens, well below the 1,048,576-token maximum.

Self-hosting moves patching, access control and logging to the operator. See the Self-Hosted LLM Security Guide.

Deprecation status

As of 15 November 2026, Meta's documentation, models page and model card list no deprecation, end-of-support date or replacement for Muse Spark Open. Downloaded weights do not expire; hosted providers may retire their endpoints on their own timelines. Monitor Meta's developer documentation rather than relying on this snapshot.

Model or itemDetailsSource
Muse-Spark-Open-816B-A42B (BF16, NVFP4)No deprecation listed (2026-11-15)Model card; models page
Hosted provider endpointsSet by each provider; not stated by MetaDocumentation

How Muse Spark Open differs from other models in its family

As of 2026-11-15. Source: Meta models page, Muse Spark Open documentation and model card. Models listed alphabetically.

ModelAPI IDWeightsContextInputPositioning (Meta wording)
Muse Embermuse-ember-1.0ClosedNot stated on the models page (2026-11-15)Text, image, video, audio, PDFEfficient model for high-volume work such as sub-agent tasks and tool calls
Muse GlimmerNoneOpenNot stated on the models page (2026-11-15)Text, imageOpen-weight, self-hosted, distilled from Muse Spark
Muse Sparkmuse-spark-1.3, 1.2, 1.1Closed1,048,576Text, image, video, audio, PDFAgentic and coding work at scale
Muse Spark OpenNoneOpen (gated)1,048,576Text, image, video, audioOpen-weight, self-hosted or provider-hosted
Muse Wondermuse-wonder-1.0Closed1,048,576Text, image, video, audio, PDFComplex software engineering and long-horizon agentic work; in preview

On Meta's models page, the open model lists text, image, video and audio input, while the hosted Muse models also list PDF. Its reasoning effort runs from low to xhigh and cannot be switched off. For other open-weight models, see Open-Weight Frontier Models Compared.

Practitioner checklist

  • Have legal review the Muse Model License, the Usage Policy and the "Built with Muse Spark Open" attribution duty before any product use.
  • If you resell inference, track gross revenue against the US $20 million condition that triggers a separate agreement with Meta.
  • Decide who in the organisation completes the gated form, since the details cannot be edited after submission.
  • Size hardware from Meta's figures: about 700 GB VRAM for NVFP4 and roughly 2 TB for BF16, plus headroom for long contexts.
  • Download only from Meta or the official Hugging Face repositories, keep original file names and record your own file hashes.
  • Bind self-hosted inference servers to localhost or a private network and require authentication on every endpoint.
  • For hosted use, review each provider's retention, training-use and region terms, as Meta does not set them.
  • Add your own guardrails and require human confirmation for irreversible tool actions, as the model card advises.
  • Store provider API keys in a secrets manager and set spend limits. See the LLMjacking Guide.
  • Recheck Meta's documentation for the full evaluation report and any revised weights.

Standards and references

Related NHI Mgmt Group resources: Meta Models, Meta Muse Spark, Open-Weight Frontier Models Compared, Open-Weight Model Security Guide, Self-Hosted LLM Security Guide.

PDF for subscribers

Take this guide with you

Every section, table and checklist in one PDF, with links to the sources and related guides. Subscribe to our weekly NHI & AI Identity Journal and we’ll email you a download link.

Free weekly newsletter

Subscribe to the NHI & AI Identity Journal

The latest on NHI and Agentic AI security – articles, research, breaches, news and events every week.

Bonus 33% off our NHI Course when you subscribe.

NHIMG Editorial Note
Researched and written by the NHI Mgmt Group editorial team using AI tools, under the editorial direction of Lalit Choda. Facts are drawn from the sources listed on this page and checked by our automated quality process. Last updated 15 November 2026. Spotted an error? Use our contact page.
NHI Mgmt Group, the independent authority on Non-Human Identity, IAM and Agentic AI security. nhimg.org

PDF Downloads for Newsletter Subscribers

Get the guide

Enter your details and we’ll email you a download link.

The full guide is free to read online. The PDF is a thank you for subscribing.

Fill in all fields marked * and tick the PDF box to continue.