--- title: "Meta Muse Spark: the free model most people haven't tried" canonical_url: "https://tryiro.com/blog/meta-muse-spark" site: "Iro AI" site_url: "https://tryiro.com" app_store: "https://apps.apple.com/app/id6759628066" language: en-US keywords: ["Meta Muse Spark", "Muse Spark benchmarks", "Meta Superintelligence Labs", "Meta AI model", "free AI model", "Muse Spark vs Llama"] date_published: "2026-07-26" date_modified: "2026-07-26" reading_time_minutes: 6 author: "Alex Furukawa" license: "© 2026 Iro AI" canonical_llm_reference: "https://tryiro.com/llms-full.txt" pillar: "ai-tools" --- # Meta Muse Spark: the free model most people haven't tried > Meta walked away from Llama and rebuilt from scratch. Muse Spark is the result — a free, multimodal reasoning model that quietly leads on medical benchmarks. **Canonical:** https://tryiro.com/blog/meta-muse-spark **Published:** 2026-07-26 **Reading time:** ~6 min **Author:** Alex Furukawa — Founder of Iro AI ## Key takeaways - Muse Spark is the first model from Meta Superintelligence Labs, launched April 2026 and updated to version 1.1 on 9 July 2026 with a focus on multimodal reasoning for agentic tasks. - It is not a Llama iteration: internally codenamed 'Avocado', it was a ground-up rebuild of Meta's architecture, infrastructure, and data pipelines. - It is free, which makes it the most accessible capable model on the market — the usual reason to skip a model (cost) does not apply. - It scored 52 on the Artificial Analysis Intelligence Index v4.0 at launch — fourth overall — and reportedly leads every competitor on health and medical benchmarks. ## The short answer **Muse Spark is Meta's first model from its Superintelligence Labs division, launched in April 2026 and updated to 1.1 on 9 July 2026 — and unlike every other model in this class, it is free.** It scored 52 on the Artificial Analysis Intelligence Index v4.0 at launch, placing fourth overall, and reportedly leads all competitors on health and medical benchmarks. The strategic story is bigger than the benchmark. Muse Spark is not a Llama update — it was built from the ground up under a new organisation, which is Meta signalling that its previous approach was not going to reach the frontier. ## Why Meta left Llama behind Meta spent years as the standard-bearer for open-weight models. Llama made capable AI available to anyone who could run it, and Llama fine-tunes repeatedly showed that the open community could match closed models on specialized tasks at a fraction of the cost. But Llama was not reaching the absolute frontier, and Muse Spark is the response: a proprietary model, internally codenamed 'Avocado', built by Meta Superintelligence Labs — a division formed under Alexandr Wang following Meta's multi-billion-dollar investment in Scale AI. It is described as a ground-up rebuild of Meta's entire AI stack: new architecture, new infrastructure, new data pipelines. The efficiency claim is the interesting part. Muse Spark reportedly achieves its reasoning capability using over an order of magnitude less compute than Llama 4 Maverick. If that holds, it explains the free pricing: a model that costs dramatically less to serve can be given away as a distribution play in a way a frontier model cannot. ## What it's actually good at Muse Spark reportedly leads every competitor on health and medical AI benchmarks. That is a genuinely differentiated position — most labs treat medical performance as one category among many, and a model that is best-in-class there has a real use case even if it sits fourth overall. The obvious caveat applies with force: strong medical benchmark performance is not medical advice, and a model being good at exam-style medical questions is not the same as being safe to rely on for a real health decision. Verify anything consequential with an actual clinician. Muse Spark introduced a multi-agent 'Contemplating' mode that reportedly beats both GPT-5.4 and Gemini on Humanity's Last Exam, one of the harder reasoning evaluations in circulation. Multi-agent approaches — several model instances working a problem and reconciling answers — are one of the few genuinely new architectural ideas to reach production in 2026. The 1.1 update, released 9 July 2026 on the Meta Model API, targets multimodal reasoning for agentic tasks specifically. That puts Muse Spark in the same race as everyone else this summer: the fight is over models that can look at things and take actions, not models that can chat. ## What 'free' really means here Free removes the main reason people don't try a second model. If you have been meaning to compare outputs across models — a genuinely useful habit, and one of the fastest ways to develop judgment about AI — Muse Spark costs nothing to add to that comparison. Two things worth being clear-eyed about: - **Free models are distribution strategies.** Meta is buying adoption and, in the usual pattern, data and ecosystem lock-in. That is not sinister, but it should inform what you paste into it. Read the terms if you handle anything sensitive. - **Fourth place is still fourth place.** On the hardest reasoning tasks, the paid frontier models — GPT-5.6 Sol, Claude Fable 5, Claude Opus 5 — remain ahead. Muse Spark is an excellent free option, not a free replacement for the top tier. The best use of a free capable model is as a second opinion. Run the same prompt through it and your usual assistant, and where they disagree, you have found the part of the answer that deserves your attention. That habit — treating model disagreement as a signal — is worth more than any single model choice. ## FAQ **What is Meta Muse Spark?** Muse Spark is the first AI model from Meta Superintelligence Labs, launched in April 2026 and updated to version 1.1 on 9 July 2026. Internally codenamed 'Avocado', it is a ground-up rebuild rather than a Llama iteration, and it is free to use. **Is Muse Spark better than Llama?** It is a different lineage, not an upgrade. Muse Spark is a proprietary ground-up rebuild of Meta's AI stack — new architecture, infrastructure, and data pipelines — and reportedly reaches its reasoning capability using over an order of magnitude less compute than Llama 4 Maverick. **Is Meta Muse Spark free?** Yes. That is its main practical advantage over comparable models — GPT-5.6, Claude Opus 5, and Grok 4.5 all charge per token, while Muse Spark does not, making it the easiest capable model to add as a second opinion. **What is Muse Spark's Contemplating mode?** A multi-agent reasoning mode in which several model instances work a problem before reconciling an answer. Meta reports it outperforming GPT-5.4 and Gemini on Humanity's Last Exam, a difficult reasoning evaluation. **How does Muse Spark rank against other AI models?** It scored 52 on the Artificial Analysis Intelligence Index v4.0 at launch, placing fourth overall. It reportedly leads all competitors specifically on health and medical benchmarks, though top-tier paid models remain ahead on general reasoning. ## Read next - [Every 2026 AI model release, tracked](https://tryiro.com/blog/new-ai-models-2026) - [The best free AI tools](https://tryiro.com/blog/best-free-ai-tools-2026) - [Claude Opus 5: what changed](https://tryiro.com/blog/claude-opus-5) ## About the author Alex Furukawa — Founder of Iro AI. Alex Furukawa is the founder of Iro AI, the gamified app for learning to use AI well. He writes about practical AI fluency — prompting, AI tools, and the daily habits that turn AI from a novelty into real leverage.