Muse Spark is a proprietary, natively multimodal reasoning model from Meta Superintelligence Labs, first released on 8 April 2026 and now offered in three versions (1.1, 1.2 and 1.3) on the Meta Model API. This profile collects the published facts about it in one place: API identifiers, context window, modalities, reasoning effort settings, Standard and Contributor mode prices, rate limits, availability, the Safety and Preparedness Report and deprecation status. It is written for developers, architects, buyers and governance teams who need checked numbers rather than launch coverage. Every fact comes from Meta's own developer documentation, announcements and safety report, 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 documentation as of November 2026. Model details change, so confirm against Meta's current documentation.
Key takeaways
- Three versions on the Model API. Meta lists muse-spark-1.3, muse-spark-1.2 and muse-spark-1.1, and marks 1.3 as "Recommended for new work".
- 1,048,576-token context for every version. Input is text, image, video, audio and PDF; output is text. Meta does not state a maximum output or knowledge cutoff on the pages checked.
- $1.25 input and $4.25 output per million tokens. Standard mode, with cached input at $0.15 and no long-context premium. Contributor mode costs $0.10 and $0.20 in exchange for permission to train future Meta models on your content.
- Moderate or lower residual risk in Meta's safety report. The Safety and Preparedness Report of 26 May 2026 rates the unmitigated model high risk for chemical and biological threats, and moderate or lower after safeguards in all measured domains.
- No deprecation notice found. Meta's Muse Spark page lists no retirement date for any version on the pages checked (2026-11-15).
What is Muse Spark?
Muse Spark is the first model in Meta's Muse family. 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, and calls it the first step on Meta's "scaling ladder". At launch it powered meta.ai and the Meta AI app, with a private API preview for select users. Meta also states that pre-training reached the same capability levels as Llama 4 Maverick with more than an order of magnitude less compute.
On 9 July 2026 Meta released Muse Spark 1.1, which it calls a "significant upgrade", together with the public preview of the Meta Model API. The 1.1 announcement lists gains in tool and computer use, coding and multimodal understanding, a 1M-token context window that the model can actively manage, and training to act as either a main agent or a parallel subagent. Meta's developer documentation now positions Muse Spark between the lower-cost Muse Ember and the capability-first Muse Wonder: it is "built for agentic and coding work at scale". Muse Spark is closed weight; Meta does not state a parameter count. The open-weight Muse Spark Open is a separate release with its own profile. For the full Meta line-up, see Meta Models.
Key facts at a glance
As of 2026-11-15. Source: Meta Muse Spark documentation, models page, pricing page and announcements.
| Field | Value | Source |
|---|---|---|
| Provider | Meta (Meta Superintelligence Labs) | Launch post; Muse Spark page |
| API IDs | muse-spark-1.3 (recommended), muse-spark-1.2, muse-spark-1.1; Contributor variants muse-spark-1.3-contributor and muse-spark-1.2-contributor | Muse Spark page; pricing page |
| Release date | 2026-04-08 (Muse Spark); 2026-07-09 (1.1); 1.2 and 1.3 not stated by Meta on the pages checked (2026-11-15) | Launch post; Model API post |
| Status | Current; 1.3 recommended for new work | Muse Spark page |
| Context window | 1,048,576 tokens (all versions) | Muse Spark page; models page |
| Max output | Not stated by Meta on the pages checked (2026-11-15) | Muse Spark page |
| Knowledge cutoff | Not stated by Meta on the pages checked (2026-11-15) | Muse Spark page |
| Input modalities | Text, image, video, audio, PDF | Muse Spark page |
| Output modalities | Text | Muse Spark page |
| Reasoning effort | minimal to max on muse-spark-1.3, default high | Muse Spark page; choose a model page |
| Endpoints | Responses, Chat Completions and Messages | Muse Spark page |
| Open weights | No | Models page |
| Licence | Proprietary | Models page |
| Parameters | Not stated by Meta on the pages checked (2026-11-15) | Models page |
| Risk rating | Moderate or lower residual risk in all measured frontier domains, after safeguards | Safety and Preparedness Report |
Release and version history
As of 2026-11-15. Dates from Meta announcements and the Safety and Preparedness Report.
| Date | Event | Source |
|---|---|---|
| 2026-04-08 | Muse Spark introduced on meta.ai and the Meta AI app; private API preview for select users; Contemplating mode begins gradual rollout on meta.ai | Launch post |
| 2026-05-26 | Muse Spark Safety and Preparedness Report published | Safety and Preparedness Report |
| 2026-07-09 | Muse Spark 1.1 released with the Meta Model API public preview; available in Thinking mode in the Meta AI app and on meta.ai | Model API post |
| Not stated | Muse Spark 1.2 (now "the previous version") and Muse Spark 1.3 ("the latest version, tuned for agentic workflows and improved coding over 1.2") | Muse Spark page |
Meta's Muse Spark page describes 1.1 as "the original version" on the Model API. Meta does not publish release dates for 1.2 or 1.3 on the documentation pages checked; dates for those versions found only in press coverage are left out here. Meta states that the models list endpoint of the Model API returns the live catalogue with created timestamps.
How much does Muse Spark cost?
As of 2026-11-15. Price per million tokens, pay-as-you-go. Meta's pricing page shows amounts with a dollar sign and does not name a currency. One price table applies to all Muse Spark versions. Source: Meta Model API pricing and rate limits.
| Item | Price | Notes | Source |
|---|---|---|---|
| Standard mode | Input: $1.25; Cached input: $0.15; Output: $4.25 | muse-spark-1.3, 1.2 and 1.1; content not used to train Meta models | Meta pricing page |
| Contributor mode | Input: $0.10; Cached input: $0.002; Output: $0.20 | muse-spark-1.3-contributor and muse-spark-1.2-contributor only; content may be used to train future Meta models | Meta pricing page |
| Web search grounding | $2.50 per 1,000 search queries | Billed on top of token costs for the web_search tool | Meta pricing page |
| Long-context premium | None | Same rate at any context length | Meta pricing page; choose a model page |
| Batch | Not stated by Meta on the pages checked (2026-11-15) | None | Meta pricing page |
- Contributor mode terms: the pricing page says Contributor mode "runs the same models at heavily discounted token prices, in exchange for permission to use your content to train future Meta models", and defines that content as "the prompts, code, files, and context you send, plus the model responses you get back".
- Muse Code default: Muse Code subscriptions default to Contributor mode, which users can turn on or off in settings.
- Caching: cached tokens are billed at the cached-input rate; the cached_tokens field in responses shows how many tokens were served from cache.
Worked example at Standard rates: a request with 100,000 input tokens and 10,000 output tokens costs $0.125 for input plus $0.0425 for output, $0.1675 in total, before search fees. The same request in Contributor mode costs $0.012.
Context window and output limits
All three versions share a 1,048,576-token context window. Because there is no long-context premium, the price per token does not change with prompt length, unlike Muse Ember, which charges higher rates above 200K input tokens. Rate limits apply per team rather than per API key, and each model ID has its own limits.
| Limit | Value | Notes | Source |
|---|---|---|---|
| Context window | 1,048,576 tokens | All versions | Muse Spark page; models page |
| Max output | Not stated by Meta on the pages checked (2026-11-15) | None | Muse Spark page |
| Rate limit, muse-spark-1.3, 1.2, 1.1 | 3,000 RPM, 4,000,000 TPM | Per team, per model | Meta pricing page |
| Rate limit, Contributor variants (1.3 and 1.2) | 150 RPM, 3,000,000 TPM | Per team, per model | Meta pricing page |
| Long-context pricing threshold | None | No long-context premium on Muse Spark | Meta pricing page |
Modalities and tools
Muse Spark accepts text, image, video, audio and PDF input and returns text. Meta's Muse Spark page lists image, video and audio understanding, tool calling, structured output and search grounding as capabilities, each with its own guide.
- Reasoning effort: on muse-spark-1.3 the range is minimal to max, with high as the default. Meta's choose a model page states that max is not available on muse-spark-1.2, muse-spark-1.1 or muse-spark-1.3-contributor.
- Audio caveat: Meta warns that "Audio understanding in Muse Spark 1.3 is not fully supported" and recommends Muse Spark 1.2 or Muse Voice Transcribe for audio.
- API formats: Responses, Chat Completions and Messages endpoints. The choose a model page also gives guidance for mapping coding tools to muse-spark-1.3.
- Consumer modes: Contemplating mode orchestrates multiple agents that reason in parallel and was rolling out gradually on meta.ai from April 2026; Muse Spark 1.1 runs in Thinking mode in the Meta AI app and on meta.ai.
- Agentic use: the 1.1 announcement states the model can write scripts where automation is faster, click through interfaces where that is simpler and batch actions at each step. Meta's computer use guide uses muse-wonder-1.0, and the Muse Spark page does not list computer use among Muse Spark's capabilities.
Tool calling and search grounding give the model reach into other systems and data. Treat each connected tool and each API key as a credentialed integration with its own scope. See the LLMjacking Guide.
Where is Muse Spark available?
As of 2026-11-15. Source: Meta Muse Spark page and announcements.
| Channel | Details | Source |
|---|---|---|
| Meta Model API | Standard mode: muse-spark-1.3, 1.2, 1.1; Contributor mode: 1.3 and 1.2 | Muse Spark page |
| Muse Code | 1.3, 1.3 Contributor, 1.2 and 1.2 Contributor on pay-as-you-go and every plan; 1.1 not available | Muse Spark page |
| meta.ai and Meta AI app | Muse Spark since 2026-04-08; 1.1 in Thinking mode since 2026-07-09 | Launch post; Model API post |
| Cloud marketplaces | Not stated by Meta on the pages checked (2026-11-15) | Muse Spark page |
| Regions | Not stated by Meta on the pages checked (2026-11-15) | Muse Spark page |
Access restrictions
Muse Spark is not a restricted-access model on the Model API: Meta lists published per-model rate limits and pay-as-you-go pricing for all three versions. The Model API is described as a public preview from 9 July 2026. Before that, API access was a private preview for select users. The Safety and Preparedness Report describes a staged, percentage-limited rollout before broad availability, along with mechanisms to deter persistent malicious use and monitor long-term user patterns, and offline cyber misuse classifiers as a compensating control. Contributor variants carry lower rate limits (150 RPM) than Standard mode.
Benchmarks (provider-reported, dated)
As of 2026-11-15. All figures are provider-reported by Meta, from the launch post (8 April 2026) or the Safety and Preparedness Report (26 May 2026). They describe the original Muse Spark, not a specific later version. Independent leaderboard results are not included because none were checked against their own published pages for this profile.
| Evaluation | Muse Spark | Comparator in source | Source and date |
|---|---|---|---|
| Humanity's Last Exam (Contemplating mode) | 58% | Not stated in text | Launch post, 2026-04-08 |
| FrontierScience Research (Contemplating mode) | 38% | Not stated in text | Launch post, 2026-04-08 |
| Cybench | 65.4% | Claude Opus 4.6 93.0% | Safety report, 2026-05-26 |
| CyberGym | 43.5% | Not stated in summary | Safety report, 2026-05-26 |
| Cyber misuse attack success rate | 9.0% | Peers 31.7% to 59.0% | Safety report, 2026-05-26 |
| StrongREJECT v2 multi-turn attack success | 44.6% | Claude Opus 4.6 31.7%, GPT-5.4 37.6% | Safety report, 2026-05-26 |
| AgentDojo prompt injection (pass@1) | 11.7% | Not stated in summary | Safety report, 2026-05-26 |
| MASK honesty | 89.1% | Not stated in summary | Safety report, 2026-05-26 |
| Severe cyber misuse refusal | 99.6% | Not stated in summary | Safety report, 2026-05-26 |
For attack success rates, lower is better; for refusal, honesty and capability scores, higher is better. Meta's 1.1 announcement links an evaluation report and shows charts, but its text gives no figures, so no 1.1, 1.2 or 1.3 scores appear here.
Model card and system card
Meta does not link a separate model card for Muse Spark on the pages checked. The closest equivalent is the Muse Spark Safety and Preparedness Report (26 May 2026), with an evaluation methodology document, both published under Meta's Advanced AI Scaling Framework. What the report states:
- Chemical and biological: the unmitigated model was rated high risk; with safeguards Meta rated residual risk moderate or lower and approved release in Meta AI. The report notes the bottleneck evaluations ran without tools.
- Cybersecurity: moderate or lower. The model solved 0 of 10 end-to-end multi-host attack chains and about 70% of isolated atomic challenges.
- Loss of control: moderate or lower, with no sufficient evidence of sandbagging, though Meta noted some self-preservation and shutdown-resistance tendencies in controlled settings that warrant monitoring.
- Evaluation awareness: Apollo Research reported the highest rate of evaluation awareness it has observed. Meta's follow-up found a behavioural effect in 3 of 20 evaluations, none tied to dangerous capabilities or launch decisions.
- Known weaknesses: the report lists susceptibility to adaptive jailbreaks and prompt injection in agentic settings, elevated sycophancy (62.9%) and higher agentic misuse on AgentHarm (26.0%) as areas of ongoing work.
For how labs structure these frameworks, see the Frontier AI Model Risk Frameworks Guide.
Data handling
Meta's Muse Spark page ties data use to the model ID you choose on pay-as-you-go.
- Standard mode: your content is "not used to train Meta models".
- Contributor mode: your content "may be used to train future Meta models", covering prompts, code, files, context and model responses.
- Switching: the mode is set by the model ID (for example muse-spark-1.3 or muse-spark-1.3-contributor), so a configuration change can move traffic between data regimes.
- Retention, residency and zero data retention: not stated by Meta on the pages checked (2026-11-15).
Because the training regime is selected by a model string, control which IDs each API key may call and alert on any use of -contributor IDs where your policy forbids it. See How to find LLM API keys in code and CI.
Deprecation status
As of 15 November 2026, Meta's Muse Spark, models and pricing pages list no deprecation notice, shutdown date or replacement for any Muse Spark version. Muse Spark 1.1 remains on the Model API in Standard mode but is not offered in Muse Code. Monitor Meta's developer documentation rather than relying on this snapshot.
| Model or item | Details | Source |
|---|---|---|
| muse-spark-1.3 and -contributor | No deprecation listed (2026-11-15); recommended for new work | Muse Spark page |
| muse-spark-1.2 and -contributor | No deprecation listed (2026-11-15) | Muse Spark page |
| muse-spark-1.1 | No deprecation listed (2026-11-15); Standard mode only; not in Muse Code | Muse Spark page |
How Muse Spark differs from other models in its family
As of 2026-11-15. Source: Meta models, choose a model and pricing pages. Models listed alphabetically.
| Model | API ID | Input / output per 1M | Context | Contributor mode | Positioning (Meta wording) |
|---|---|---|---|---|---|
| Muse Ember | muse-ember-1.0 | $0.10 / $0.50 up to 200K input; higher above | Not stated by Meta on the pages checked (2026-11-15) | No | Lowest-priced Muse API model |
| Muse Spark | muse-spark-1.3, 1.2, 1.1 | $1.25 / $4.25 | 1,048,576 | Yes (1.3 and 1.2) | Built for agentic and coding work at scale |
| Muse Spark Open | No Model API ID (open weights) | Self-hosted | Not stated by Meta on the pages checked (2026-11-15) | Not applicable | 816B total, 42B active mixture-of-experts |
| Muse Wonder | muse-wonder-1.0 | $4.00 / $20.00 | 1,048,576 | No | Capability-first work such as multi-file changes and multi-hour agent runs |
The choose a model page frames Muse Spark as the lower-cost main loop for agents and assistants compared with Muse Wonder, and recommends Spark or Wonder over Ember for hard single-shot coding, knowledge-heavy questions and visual coding. Only Muse Spark has Contributor variants on the Model API. See Meta Muse Spark Open for the open-weight model.
Practitioner checklist
- Pin a versioned ID such as muse-spark-1.3 in code and record the date you adopted it.
- Decide Standard or Contributor mode as a data policy decision, not a pricing one, and document who approved it.
- Block or alert on -contributor model IDs for keys that handle confidential, personal or regulated data.
- Check Muse Code settings, where subscriptions default to Contributor mode.
- Route audio workloads to muse-spark-1.2 or Muse Voice Transcribe while Meta flags 1.3 audio as not fully supported.
- Plan for 150 RPM on Contributor variants and 3,000 RPM on Standard, applied per team.
- Budget search grounding at $2.50 per 1,000 queries on top of token costs.
- Store Model API keys in a secrets manager, scope them per environment and rotate them on staff changes.
- Recheck Meta's Muse Spark and pricing pages monthly for new versions or retirements.
Standards and references
- Meta: Muse Spark documentation (accessed 2026-11-15)
- Meta: Model API pricing and rate limits (accessed 2026-11-15)
- Meta: Model API models (accessed 2026-11-15)
- Meta: Choose a model (accessed 2026-11-15)
- Meta: Muse Spark Safety and Preparedness Report (accessed 2026-11-15)
- Meta AI: Introducing Muse Spark (accessed 2026-11-15)
- Meta AI: Muse Spark 1.1 and the Meta Model API (accessed 2026-11-15)
Related NHI Mgmt Group resources: Meta Models, Meta Muse Spark Open, Frontier AI Models Compared, LLMjacking Guide, How to set spend limits on LLM API keys.