Muse Spark is Meta's proprietary large language model, built by its Meta Superintelligence Labs (MSL) division. It drives the Meta AI assistant across WhatsApp, Instagram, Messenger and Ray-Ban Meta glasses, where it costs nothing to use. Developers pay per token through the Meta Model API, starting at $0.10 per million tokens on the Contributor tier. The model reasons over text, images, audio and PDFs, with a context window of one million tokens.
Muse Spark reads text and images natively, with a visual chain of thought that can annotate what it sees. You snap a photo of an airport snack shelf and it ranks the packets by protein content. On health topics, Meta had its training data reviewed by over 1,000 physicians, so the model can break down the nutrition of a meal or the muscles worked by an exercise.
This large language model also marks a strategic pivot. After years of freely released Llama weights, Meta is keeping this one behind closed doors. The MSL lab, built around Alexandr Wang, rebuilt its entire training stack in nine months before shipping this first model, long known under the codename Avocado (really).
The Muse family grew fast, with three releases in four months. Version 1.1 opened the Meta Model API and added a Thinking mode, while 1.2 doubled down on programming and arrived alongside Muse Code, a terminal coding agent trained hand in hand with the model.
On independent measurements, Artificial Analysis scored it 57 in early August, twelfth out of 185 models tested, a figure worth checking against live AI model rankings since positions shift weekly. It did not dethrone Claude Opus 5 or GPT-5.6, but neither of those comes free at this scale.
| Version | Released | Highlights |
|---|---|---|
| Muse Spark | April 2026 | First MSL model, the engine behind Meta AI |
| Muse Spark 1.1 | July 2026 | Agentic focus, Meta Model API public preview |
| Muse Spark 1.2 | August 2026 | Coding update, shipped with Muse Code, listed on OpenRouter |
Consumer access is free: Meta AI, powered by Muse Spark, answers at no charge across Meta's apps and on meta.ai. For developers, two API tiers coexist. Standard bills $1.25 per million input tokens and $4.25 per million output tokens, with a commitment that your data stays out of training; Contributor drops to $0.10 and $0.20 in exchange for letting Meta use your prompts. New accounts get $20 in credits. Meta adjusts these numbers as the family grows, so treat the developer page as the final word.
One practical caveat: the API preview opened to US developers first, while the consumer assistant rolled out much more widely from day one.
Yes, for everyday use: Muse Spark answers at no cost in the Meta AI app, on meta.ai and inside Meta's messaging apps. API access is pay-as-you-go instead, with $20 in welcome credits when you sign up. No monthly subscription exists so far.
Independent benchmarks place Muse Spark near the top of the global table, behind the latest OpenAI and Anthropic releases but ahead of most freely accessible models. Its strengths are health questions, chart and image reading, and API rates roughly a quarter of what its biggest rivals charge.
It depends on the tier. Standard API traffic is not reused to improve Meta's products, while the far cheaper Contributor tier explicitly grants Meta that right. Keep in mind that signing in requires a Facebook or Instagram account, a detail that fuels most privacy discussions around the model.
Muse Spark is the model, and Muse Code is the terminal-based coding agent built on top of it. Trained together, they form Meta's software development duo: the agent writes, debugs and runs parallel subagents on long projects, competing on the same ground as Anthropic's Claude Code.
Verdict: A top-15 model for zero dollars is the pitch: anyone already living in WhatsApp or Instagram can call on Muse Spark today, and cost-conscious development teams get some of the lowest per-token rates on the market.
