Meta Drops Muse Spark — Free, Closed-Source, and Coming for GPT and Claude
TL;DR
Meta launched Muse Spark, the first model from Meta Superintelligence Labs. It’s closed-source (a major break from Llama’s open approach), free for 3B+ users, built from scratch with native multimodality, and reaches frontier-class performance with 10x less compute than Llama 4.
What Happened
Meta released Muse Spark on April 8, 2026 — built under the leadership of Alexandr Wang, who joined Meta through a $14.3B deal with Scale AI.
This isn’t a Llama variant. It’s a completely new architecture built from the ground up. Native multimodal input (text, images, voice), built-in tool use, visual chain-of-thought reasoning, and multi-agent orchestration — all native, not bolted on.
The biggest surprise: it’s closed-source. Meta, the company that built its AI reputation on open-source Llama, is keeping this one proprietary. They’ve said future versions may be open-sourced, but Muse Spark itself stays closed.
The Numbers
| Benchmark | Muse Spark | GPT-5.4 | Claude Opus 4.6 | Gemini 3.1 Pro |
|---|---|---|---|---|
| AI Intelligence Index | 52 | — | — | — |
| MMMU-Pro (multimodal) | 80.5% | — | — | 82.4% |
| CharXiv Reasoning | 86.4 | 82.8 | — | 80.2 |
| HealthBench Hard | 42.8 | 40.1 | — | — |
| HLE (Contemplating) | 50.2% | 43.9% | — | 44.7% |
Where it leads: Healthcare benchmarks, chart/data comprehension, scientific reasoning, and token efficiency.
Where it trails: Coding (Terminal-Bench: 59.0 vs GPT-5.4’s 75.1), abstract reasoning (ARC-AGI-2: 42.5 vs 76.1), and agentic tasks.
Three Reasoning Modes
Muse Spark doesn’t just have one way of answering — it has three:
Instant — Single response, low latency. Quick questions, fast answers.
Thinking — Multi-step analysis. Reasons step by step before answering.
Contemplating — The interesting one. Launches parallel sub-agents that work on the problem simultaneously, then synthesizes the answers.
Where You Can Use It
Rolling out across:
- Meta AI app and website (meta.ai) — live now
- WhatsApp — coming weeks
- Instagram — coming weeks
- Facebook — coming weeks
- Messenger — coming weeks
- Ray-Ban Meta AI glasses — coming weeks
All completely free. No subscription, no API key. No public API yet — private preview only for select partners.
The Strategic Shift
This is Meta’s biggest AI pivot since acquiring Alexandr Wang’s involvement:
From open to closed. Llama was open-source. Muse Spark is not. Meta is betting that a proprietary, tightly integrated consumer AI beats open-weight models for their 3B+ user base.
From big to efficient. Muse Spark reaches Llama 4 Maverick performance with 10x less compute. The entire benchmark suite ran on just 58M output tokens — compared to 157M for Claude Opus 4.6.
From chat to platform. Muse Spark isn’t just a chatbot. It’s designed to surface content from across Meta’s platforms, understand your world through the data you already generate, and act as a persistent AI assistant embedded in every Meta product.
The Investment
Meta is spending $115-135 billion on AI infrastructure in 2026 — nearly double last year’s capex. This isn’t experimentation. This is an all-in bet.
Why It Matters
Three billion people now have access to a frontier-class AI model for free. No subscription, no API key, no setup — just open WhatsApp and ask.
That changes the competitive landscape. OpenAI charges $20-200/month. Anthropic charges $20/month. Google gives Gemini away but with limits. Meta is giving away frontier performance to everyone, everywhere, subsidized by ad revenue.
The question isn’t whether Muse Spark is the best model. It’s whether “free and everywhere” beats “best and paid.”
Follow Synvoya for daily AI news summaries — quick reads, no fluff.