Meta now offers two very different model families side by side: Muse Spark, a closed-weight reasoning model sold through the Meta Model API, and Llama 4, the open-weight family released in April 2025 that Muse Spark followed. This page compares them on the same fields: model IDs, release dates, status, context window, modalities, parameters, prices, licence, availability, provider-reported benchmarks, safety documentation and lifecycle. It is written for developers deciding between a hosted API and self-hosted weights, buyers comparing cost models and governance teams who need to know which licence and data terms apply. Models are listed by release date, newest first, an objective order that implies no ranking; the two Llama 4 models share a release date and are listed alphabetically. Every fact comes from Meta's own developer documentation, announcements, model cards and safety report, cited in the sources list.
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 developer documentation as of November 2026. Model details change, so confirm against Meta's current documentation.
Key takeaways
- Closed API model versus open weights. Meta's models page lists every Muse Spark version with open weights "No"; Llama 4 Scout and Maverick weights are downloadable under the Llama 4 Community License Agreement.
- Muse Spark is priced per token; Llama 4 is not sold by Meta per token. Muse Spark Standard mode costs $1.25 input and $4.25 output per million tokens. Meta gives Llama 4 Maverick only a serving cost estimate of $0.19 per million tokens (3:1 blended) with distributed inference.
- Context: 1M for Muse Spark and Maverick, 10M for Scout. Every Muse Spark version lists 1,048,576 tokens; Meta's Llama 4 model card lists 10M tokens for Scout and 1M for Maverick.
- Meta links the two directly. Its Muse Spark launch post states that Meta can "reach the same capabilities with over an order of magnitude less compute than our previous model, Llama 4 Maverick".
- Neither has a published retirement date. Meta's Muse Spark pages list no deprecation for any version, and the Llama 4 Scout card calls it "a static model trained on an offline dataset".
What are Muse Spark and Llama 4?
Muse Spark is the first model in Meta's Muse family from Meta Superintelligence Labs. Meta's launch post of 8 April 2026 describes it as a natively multimodal reasoning model with tool use, visual chain of thought and multi-agent orchestration. Meta's models page calls it "Meta's model for agentic and coding work". It is offered in three versions on the Meta Model API, 1.1, 1.2 and 1.3, and Meta states no parameter count or licence for it. Deep detail is in the Meta Muse Spark profile.
Llama 4 is Meta's earlier open-weight family. The Hugging Face model card dates the release to 5 April 2025 and describes two Mixture-of-Experts models with early fusion for native multimodality: Llama 4 Scout (17B active, 109B total, 16 experts) and Llama 4 Maverick (17B active, 400B total, 128 experts), each in pretrained and instruction-tuned variants. Meta's Llama 4 page still labels the family "LATEST" among Llama releases, but Meta's current models page for the Model API lists no Llama model.
A third Meta model sits between the two: Muse Spark Open, which the models page calls "the open-weight release of Muse Spark" (816B total, 42B active). It is a separate model and is covered in the Meta Muse Spark Open profile. For the whole Meta line-up, see Meta Models.
Every model at a glance
As of 2026-11-15. Sources: Meta models page, Muse Spark page, Llama 4 page, Llama 4 model card and Hugging Face card.
| Model | Model ID | Released | Status | Open weights | Parameters | Licence |
|---|---|---|---|---|---|---|
| Muse Spark 1.3 | muse-spark-1.3; muse-spark-1.3-contributor | Not stated by Meta on the pages checked (2026-11-15) | Current, "Recommended for new work" | No | Not stated by Meta | Not stated by Meta (closed weight) |
| Muse Spark 1.2 | muse-spark-1.2; muse-spark-1.2-contributor | Not stated by Meta on the pages checked (2026-11-15) | Current, "the previous version" | No | Not stated by Meta | Not stated by Meta (closed weight) |
| Muse Spark 1.1 | muse-spark-1.1 | 2026-07-09 | Current, "the original version" on the Model API | No | Not stated by Meta | Not stated by Meta (closed weight) |
| Llama 4 Maverick | meta-llama/Llama-4-Maverick-17B-128E (and -Instruct) | 2025-04-05 | Static model; no end-of-life date stated | Yes (BF16 and FP8) | 17B active, 400B total, 128 experts | Llama 4 Community License Agreement |
| Llama 4 Scout | meta-llama/Llama-4-Scout-17B-16E (and -Instruct) | 2025-04-05 | Static model; no end-of-life date stated | Yes (BF16) | 17B active, 109B total, 16 experts | Llama 4 Community License Agreement |
The original Muse Spark was introduced on 8 April 2026 on meta.ai and the Meta AI app, with a private API preview; Muse Spark 1.1 followed on 9 July 2026 with the public preview of the Meta Model API. Meta's documentation gives no release dates for 1.2 and 1.3; dates found only in press coverage are left out.
When was each model released?
As of 2026-11-15. Dates from Meta announcements, the Llama 4 Hugging Face card and the Muse Spark Safety and Preparedness Report.
| Date | Event | Model ID |
|---|---|---|
| 2025-04-05 | Llama 4 Scout and Llama 4 Maverick released as open weights | meta-llama/Llama-4-Scout-17B-16E, meta-llama/Llama-4-Maverick-17B-128E |
| 2026-04-08 | Muse Spark introduced on meta.ai and the Meta AI app; private API preview for select users; Contemplating mode rolling out gradually | Not stated in the launch post |
| 2026-05-26 | Muse Spark Safety and Preparedness Report published under Meta's Advanced AI Scaling Framework (v2) | Not applicable |
| 2026-07-09 | Muse Spark 1.1 released with the Meta Model API public preview; in Thinking mode on meta.ai and the Meta AI app | muse-spark-1.1 |
| Not stated | Muse Spark 1.2 and 1.3 added to the Model API | muse-spark-1.2, muse-spark-1.3 |
How much does each model cost?
As of 2026-11-15. Per million tokens. Meta's pricing page shows amounts with a dollar sign and does not name a currency.
| Model | Input | Cached input | Output | Batch | Notes stated by provider |
|---|---|---|---|---|---|
| Muse Spark 1.3, 1.2, 1.1 (Standard mode) | $1.25 | $0.15 | $4.25 | Not stated by Meta on the pages checked (2026-11-15) | No long-context premium; web search $2.50 per 1,000 queries on top |
| Muse Spark 1.3 and 1.2 (Contributor mode) | $0.10 | $0.002 | $0.20 | Not stated by Meta on the pages checked (2026-11-15) | Meta may use prompts, code, files, context and responses to train future models |
| Llama 4 Maverick | No per-token price from Meta; serving cost estimate $0.19 per million tokens (3:1 blended) with distributed inference, $0.30 to $0.49 projected on a single host | Not applicable | Estimate from Meta's Llama 4 page; your own compute or a hosting provider's price applies | ||
| Llama 4 Scout | No per-token price from Meta; no serving cost estimate on the pages checked (2026-11-15) | Not applicable | Meta states single H100 GPU efficiency with int4 quantisation | ||
The cost models are not like for like. Muse Spark is a metered API price that includes Meta's hosting. Llama 4 has no Meta API price: you pay for GPUs or for a third-party host, whose prices Meta does not set. Compare total cost per task, including idle capacity for self-hosting. See Self-Hosted LLMs vs Hosted LLM APIs and the LLM Pricing Tracker.
What changed between models?
Only changes Meta documents are listed.
- Llama 4 to Muse Spark: the move from published weights to a closed-weight hosted model. Meta's launch post states that Muse Spark reaches the same capabilities "with over an order of magnitude less compute than our previous model, Llama 4 Maverick". Input broadened from text and images (Llama 4) to text, image, video, audio and PDF (Muse Spark), and Muse Spark adds a configurable reasoning effort that the Llama 4 documentation does not describe.
- Muse Spark to 1.1: Meta calls 1.1 "a significant upgrade", with gains in tool and computer use, coding and multimodal understanding, an actively managed 1 million token context window, zero-shot generalisation to new native tools, MCP servers and custom skills, and work as a main agent or parallel subagent. Meta also claims better adversarial robustness and lower hallucination and sycophancy rates.
- 1.2 to 1.3: Meta's Muse Spark page says 1.3 is "tuned for agentic workflows and improved coding over 1.2". It also warns that audio understanding in 1.3 "is not fully supported" and recommends 1.2 or Muse Voice Transcribe for audio.
See Open-Weight vs Closed Models Explained for what the change in distribution model means in practice.
What workloads does Meta position each model for?
This section records the provider's own positioning wording; it is not an assessment.
| Model | Positioning stated by provider | Related notes stated by provider |
|---|---|---|
| Muse Spark 1.3 | "Tuned for agentic workflows and improved coding over 1.2"; recommended for new work | Audio understanding not fully supported |
| Muse Spark 1.2 | The previous version | Recommended by Meta for audio understanding |
| Muse Spark 1.1 | The original version on the Model API | Standard mode only; not available in Muse Code |
| Llama 4 Maverick | Natively multimodal model for image and text understanding, with fast responses at low cost | FP8 weights fit on a single H100 DGX host |
| Llama 4 Scout | Natively multimodal model with a 10M-token context window and single H100 efficiency | Fits on one H100 with on-the-fly int4 quantisation |
Worked cost example
A workload of 10,000 requests a day, each with a 20K-token prompt and a 1K-token response, uses 200 million input and 10 million output tokens a day. On Muse Spark Standard mode that is $250 input plus $42.50 output, $292.50 a day before search fees and caching. In Contributor mode it is $20 plus $2, $22 a day, in exchange for allowing Meta to train on the content. For Llama 4 Maverick, Meta's own blended estimate of $0.19 per million tokens would put 210 million tokens at about $40 a day under distributed inference, but that is Meta's projection of serving cost, not a price anyone charges. Measure on your own prompts; the two families do not share a tokeniser on the pages checked.
What is the context window of each model?
As of 2026-11-15. From the Muse Spark page, models page, Llama 4 model card and Hugging Face card.
| Model | Context window | Max output | Knowledge cutoff | Input | Output | Reasoning controls |
|---|---|---|---|---|---|---|
| Muse Spark 1.3 | 1,048,576 tokens | Not stated by Meta on the pages checked (2026-11-15) | Not stated by Meta on the pages checked (2026-11-15) | Text, image, video, audio (not fully supported), PDF | Text | Effort minimal to max, default high; max not available on the Contributor variant |
| Muse Spark 1.2 | 1,048,576 tokens | Not stated by Meta on the pages checked (2026-11-15) | Not stated by Meta on the pages checked (2026-11-15) | Text, image, video, audio, PDF | Text | Not stated separately by Meta on the pages checked (2026-11-15) |
| Muse Spark 1.1 | 1,048,576 tokens | Not stated by Meta on the pages checked (2026-11-15) | Not stated by Meta on the pages checked (2026-11-15) | Text, image, video, audio, PDF | Text | Not stated separately by Meta on the pages checked (2026-11-15) |
| Llama 4 Maverick | 1M tokens | Not stated by Meta on the pages checked (2026-11-15) | August 2024 | Multilingual text, images (tested up to 5) | Multilingual text and code | Not stated by Meta on the pages checked (2026-11-15) |
| Llama 4 Scout | 10M tokens | Not stated by Meta on the pages checked (2026-11-15) | August 2024 | Multilingual text, images (tested up to 5) | Multilingual text and code | Not stated by Meta on the pages checked (2026-11-15) |
Llama 4 supports 12 languages per the model card (Arabic, English, French, German, Hindi, Indonesian, Italian, Portuguese, Spanish, Tagalog, Thai and Vietnamese), with image understanding in English only. A self-hosted context window is also bounded by your hardware and serving settings; Meta notes the Llama 4 context lengths were evaluated across 512 GPUs. See Context Windows Explained.
Capabilities and tools compared
As of 2026-11-15. From the Muse Spark page and the Llama 4 model card.
| Model | Function calling and structured outputs | Search and other tools | Fine-tuning | Interfaces stated by provider |
|---|---|---|---|---|
| Muse Spark 1.3, 1.2, 1.1 | Tool calling and structured output listed | Search grounding (web_search, billed per query); image, video and audio understanding | Not stated by Meta on the pages checked (2026-11-15) | Responses, Chat Completions and Messages endpoints on the Meta Model API |
| Llama 4 Maverick and Scout | Zero-shot function calling in Python-style and JSON formats, with an ipython role for tool output | Tools defined by the deployer | Weights can be modified under the licence; Meta provides pretrained and instruction-tuned variants | Self-hosted weights; prompt format in Meta's Llama 4 model card |
Images in Llama 4 are split into 336x336 tiles plus a global tile. Muse Spark's agentic features, such as subagent delegation and context compaction, are described in the 1.1 announcement.
Where is each model available?
| Model | Availability stated by provider |
|---|---|
| Muse Spark 1.3 | Meta Model API (Standard and Contributor); Muse Code on pay-as-you-go and every plan |
| Muse Spark 1.2 | Meta Model API (Standard and Contributor); Muse Code on pay-as-you-go and every plan |
| Muse Spark 1.1 | Meta Model API (Standard only); Thinking mode on meta.ai and the Meta AI app from 2026-07-09; not in Muse Code |
| Llama 4 Maverick | Meta's Llama downloads page; Hugging Face (gated) |
| Llama 4 Scout | Meta's Llama downloads page; Hugging Face (gated) |
Cloud marketplaces and regions are not stated by Meta for either family on the pages checked (2026-11-15). Third-party hosts publish their own Llama 4 IDs and prices.
Access restrictions
Muse Spark: Meta publishes pay-as-you-go prices and per-team rate limits: 3,000 RPM and 4,000,000 TPM for each Standard version, 150 RPM and 3,000,000 TPM for each Contributor variant. Limits apply per team, not per API key. The Model API has been a public preview since 9 July 2026. The safety report states that the chemical and biological capability is deployed first-party only, without an API.
Llama 4: downloads on Hugging Face are gated. The form asks for full legal name, date of birth and full organisation name with corporate identifiers, and cannot be edited after submission. The Llama 4 Community License requires a licence from Meta, at Meta's discretion, where the licensee and its affiliates had more than 700 million monthly active users in the calendar month before the release date. Distributions must include the licence and prominently display "Built with Llama", and models trained or improved with Llama outputs must start their name with "Llama". Use is also bound by Meta's Acceptable Use Policy. See the Open-Weight Model Licence Tracker.
Model cards and system cards
| Model | Model card or page | System card or safety report | Technical report |
|---|---|---|---|
| Muse Spark (1.1 to 1.3) | Muse Spark documentation page | Safety and Preparedness Report (26 May 2026, original Muse Spark) | Not stated by Meta on the pages checked (2026-11-15) |
| Llama 4 Maverick and Scout | Llama 4 model card and prompt formats; Hugging Face card | Not stated by Meta on the pages checked (2026-11-15) | Not stated by Meta on the pages checked (2026-11-15) |
The Muse Spark report rates the unmitigated model high risk for chemical and biological threats and moderate or lower after safeguards, with cybersecurity and loss of control moderate or lower. It describes a percentage-limited early rollout, detection of persistent malicious use and offline cyber misuse classifiers. Llama 4 Maverick appears in the legends of several biology figures in that report, but the text does not give its values. See Frontier AI Model Risk Frameworks Guide and Model Cards and System Cards Explained.
Benchmarks
As of 2026-11-15. Figures are provider-reported. The two families were evaluated on different suites at different times, and no benchmark appears in Meta's text for both, so no row-to-row comparison is possible.
| Model | Provider-reported benchmarks | Independent leaderboard results |
|---|---|---|
| Muse Spark (original) | Humanity's Last Exam 58% and FrontierScience Research 38%, Contemplating mode (launch post, 2026-04-08); Cybench 65.4%, CyberGym 43.5%, AgentDojo attack success 11.7%, StrongREJECT v2 attack success 44.6%, MASK honesty 89.1% (safety report, 2026-05-26) | Not checked for this page |
| Muse Spark 1.1, 1.2, 1.3 | No figures in the text of the pages checked; the 1.1 announcement shows charts only | Not checked for this page |
| Llama 4 Maverick (instruction tuned) | MMLU Pro 80.5; GPQA Diamond 69.8; LiveCodeBench (10/2024 to 02/2025) 43.4; MMMU 73.4; ChartQA 90.0; DocVQA (test) 94.4 (Hugging Face card, 2025-04) | Not checked for this page |
| Llama 4 Scout (instruction tuned) | MMLU Pro 74.3; GPQA Diamond 57.2; LiveCodeBench (10/2024 to 02/2025) 32.8; MMMU 69.4; ChartQA 88.8; DocVQA (test) 94.4 (Hugging Face card, 2025-04) | Not checked for this page |
For attack success rates lower is better; for the others higher is better. The Scout card's evaluation widget shows a different GPQA Diamond figure (46.46) from its table; the table figure is used here.
Deprecations and retirements
As of 2026-11-15.
| Model | Model ID | Notice or deprecation date | Retirement or shutdown date | Stated replacement |
|---|---|---|---|---|
| Muse Spark 1.3 | muse-spark-1.3 | None listed | None listed | None stated |
| Muse Spark 1.2 | muse-spark-1.2 | None listed | None listed | None stated |
| Muse Spark 1.1 | muse-spark-1.1 | None listed | None listed | None stated |
| Llama 4 Maverick | meta-llama/Llama-4-Maverick-17B-128E | None stated | Not applicable to downloaded weights | None stated |
| Llama 4 Scout | meta-llama/Llama-4-Scout-17B-16E | None stated | Not applicable to downloaded weights | None stated |
The lifecycle risk differs. A hosted Muse Spark version can be withdrawn by Meta, so monitor the developer documentation. Downloaded Llama 4 weights keep working, but they are static: the Scout card says future tuned versions may follow, and the knowledge cutoff stays at August 2024. See the LLM Deprecation and Retirement Tracker.
Data handling and security
Muse Spark: the training regime is chosen by model ID. Standard mode content "is not used to train Meta models"; Contributor mode content "may be used to train future Meta models". Retention, residency and zero data retention terms are not stated on the Muse Spark pages checked (2026-11-15). Because a single string switches data regimes, restrict which IDs each key may call and alert on -contributor IDs where policy forbids them. See the LLM Data Retention and Training-Use Policies page and How to find LLM API keys in code and CI.
Llama 4: prompts and outputs stay wherever you run the weights, so retention and security are the deployer's responsibility, including model file integrity, serving endpoint access and GPU host hardening. The Hugging Face access form collects identity details for gated downloads. See the Self-Hosted LLM Security Guide and the Open-Weight Model Security Guide.
How to choose a model: factual criteria
| Criterion | What the documentation shows |
|---|---|
| Deployment model | Muse Spark: Meta-hosted API only; Llama 4: self-hosted or third-party hosted weights |
| Pricing basis | Muse Spark: $1.25/$4.25 Standard, $0.10/$0.20 Contributor; Llama 4: own compute |
| Context above 1M tokens | Llama 4 Scout only (10M) |
| Video, audio and PDF input | Muse Spark only; Llama 4 takes text and images |
| Training on your content | Muse Spark Contributor mode only; Standard mode and self-hosted Llama 4 do not send content to Meta for training |
| Licence obligations | Llama 4: 700 million MAU threshold, "Built with Llama" attribution, Acceptable Use Policy; Muse Spark: API terms |
| Knowledge cutoff | Llama 4: August 2024; Muse Spark: not stated |
For other providers, see Frontier AI Models Compared and Open-Weight Frontier Models Compared.
Practitioner checklist
- Pin a versioned Muse Spark ID such as muse-spark-1.3 and record it in the AI inventory; avoid 1.3 for audio-heavy work until Meta changes its guidance.
- Block or alert on -contributor model IDs wherever policy forbids Meta training on your content.
- Plan Muse Spark capacity around per-team limits (3,000 RPM Standard, 150 RPM Contributor), not per-key limits.
- Before deploying Llama 4 commercially, check the 700 million MAU clause, attribution duties and the Acceptable Use Policy with legal counsel.
- Size Llama 4 hardware from Meta's notes: Scout on one H100 with int4, Maverick FP8 on one H100 DGX host.
- Account for the August 2024 knowledge cutoff in Llama 4 applications that need current facts.
- Store Meta Model API keys in a secrets manager and secure self-hosted Llama endpoints with authentication.
Standards and references
- Meta: Models (accessed 2026-11-15)
- Meta: Muse Spark documentation (accessed 2026-11-15)
- Meta: Pricing and rate limits (accessed 2026-11-15)
- Meta: Introducing Muse Spark (accessed 2026-11-15)
- Meta: Muse Spark 1.1 and the Meta Model API (accessed 2026-11-15)
- Meta: Muse Spark Safety and Preparedness Report (accessed 2026-11-15)
- Meta: Llama 4 (accessed 2026-11-15)
- Meta: Llama 4 model card and prompt formats (accessed 2026-11-15)
- Meta on Hugging Face: Llama-4-Scout-17B-16E-Instruct model card (accessed 2026-11-15)
Related NHI Mgmt Group resources: Meta Models, Meta Muse Spark, Meta Muse Spark Open, Open-Weight vs Closed Models Explained, LLM Pricing Tracker.