Muse Spark
Usage BasedMeta's frontier reasoning model for agents, code, and multimodal work.
Published 1 October 2026
Scores
Popularity4/5
Meta AI puts Muse Spark in front of billions of consumers, and each release is covered widely in tech press. As a developer API it only opened in July 2026, so its footprint in production codebases is still small next to OpenAI, Claude, or Gemini.
Learning Curve2/5
The Meta Model API follows the OpenAI chat-completions format, so anyone who has called a hosted LLM can switch to it by changing the base URL and model ID. Choosing between reasoning levels and the Standard or Contributor terms is the only new decision.
Flexibility3/5
Covers text, image, and video input, tool calling, and adjustable reasoning effort, but there are no weights to fine-tune or self-host and customization stops at what the API exposes.
Performance4/5
The 1.3 release posts large jumps on computer-use and knowledge-work evaluations (OSWorld 2.0 from 47.6 to 66.9) and near-perfect long-context retrieval out to 1M tokens, while Meta's own charts still place it just behind Claude Opus 5 on the hardest agentic tasks.
Portability2/5
Closed weights and a US-only first-party API tie workloads to Meta, though the OpenAI-compatible format and availability on OpenRouter keep switching costs low.
About Muse Spark
Muse Spark is the proprietary frontier model family built by Meta Superintelligence Labs (MSL), the division Meta formed after Llama 4's reception in 2025. It launched on April 8, 2026 inside Meta AI, where it replaced Llama as the model behind meta.ai, the Meta AI app, and the assistant in WhatsApp, Instagram, and Messenger. Meta describes it as a departure from its open-weight strategy: Llama remains the open-weight line, while Muse Spark is the closed flagship.
The family has moved quickly through four versions. 1.0 was consumer-only; 1.1 (July 2026) was the first release developers could call directly; 1.2 (August 2026) launched alongside the Muse Code coding agent; and 1.3 (September 2026) is the current flagship. All versions are natively multimodal reasoning models with a 1,048,576-token context window, accepting text, image, and video input and returning text. 1.3 is tuned for agentic work such as tool use, computer use, long-horizon coding, and knowledge-work tasks, and ships with selectable reasoning effort levels.
Developers access it through the Meta Model API, an OpenAI-compatible endpoint that opened as a US-only public preview in July 2026 with usage-based token pricing and starter credits. Muse Spark is also served by third-party routers such as OpenRouter and LLM Gateway, and appears in the model pickers of coding agents including OpenCode and Cline. Consumers use it for free through Meta AI and through Meta's Muse personal agent, which runs on it.
Every version also comes in a Contributor variant: the same weights at a steep discount, or free on some coding-agent platforms, in exchange for Meta using the prompts and completions to train future models. It is the usual reason Muse Spark shows up as a free option in coding tools, and it is the main privacy trade-off to check before routing production traffic to it. Contributor access carries tighter rate limits than the standard tier.
Meta has said Muse Spark 1.2's weights will be released openly, and in August 2026 it shipped Muse Glimmer, a smaller Apache-2.0 model distilled from 1.2. The current 1.3 flagship remains closed.
Key Features
- 1,048,576-token context window across all versions
- Native multimodal input: text, images, and video
- Selectable reasoning effort levels on the 1.3 flagship
- Tuned for tool use, computer use, and long-horizon agentic coding
- OpenAI-compatible Meta Model API with prompt caching
- Discounted Contributor variant of each version in exchange for training data
- Available through OpenRouter, LLM Gateway, OpenCode, and Cline
- Powers Meta AI and the Muse personal agent
Pros
- Standard token pricing sits well below the other US frontier labs for comparable agentic benchmark results
- 1M-token context with strong long-context retrieval scores on the 1.3 release
- OpenAI-compatible endpoint means existing SDKs and agent frameworks work with a base URL change
- The Contributor tier makes it free or near-free to try inside OpenCode and Cline
Cons
- The first-party Meta Model API is a US-only public preview, so teams elsewhere rely on third-party routers
- Closed weights for the current flagship, unlike Meta's own Llama and Muse Glimmer models
- Contributor pricing sends prompts and completions to Meta for training, and some tools default users to it
- Short track record as a developer API compared with OpenAI, Anthropic, and Google, with frequent version turnover
Muse Spark Pricing
Usage Based- · Free chat use through meta.ai, the Meta AI app, WhatsApp, Instagram, and Messenger
- · No API access
- · $1.25 per million input tokens, $4.25 per million output tokens
- · $0.15 per million cached input tokens
- · Same rate for versions 1.1, 1.2, and 1.3
- · Starter credits for new Meta Model API accounts
- · $0.10 per million input tokens, $0.20 per million output tokens
- · Prompts and completions may be used to train future Meta models
- · Lower rate limits than the Standard tier
Tools Related to Muse Spark
Works well with Muse Spark(1)
OpenRouter serves Muse Spark alongside other providers' models, which is how teams outside the US reach it while Meta's own Model API remains a US-only preview.
Integrates with Muse Spark(1)
Muse Spark is the model family behind Muse, Meta's personal agent, which is the main consumer product built on these models.
Alternatives to Muse Spark(5)
Muse Spark offers a 1M-token context window, video input, and a discounted Contributor tier through Meta's OpenAI-compatible API; Claude has the longer track record with developers and far wider availability, since Meta's own API is still a US-only preview.
Muse Spark's Meta Model API follows the OpenAI chat-completions format, so trying it is a base-URL change. OpenAI has the more mature developer platform and worldwide availability; Muse Spark brings a 1M-token context window and a Contributor tier that trades prompts for a lower price.
Muse Spark and Grok are the newer challengers among proprietary models, each tied to a social platform: Muse Spark powers Meta AI across WhatsApp and Instagram, while Grok draws on real-time X data. Both expose OpenAI-compatible APIs.
Muse Spark and Google Gemini are both natively multimodal models with 1M-token context windows and video input. Gemini is available worldwide through Google AI Studio and Vertex AI, while Meta's first-party API is a US-only preview reached elsewhere through routers.
Meta's two model lines: Muse Spark is the closed-weight frontier family reached only through an API, while Llama stays open-weight and can be downloaded, fine-tuned, and self-hosted. Pick Llama for control over the weights, Muse Spark for Meta's strongest model.
