In the space of about ten weeks, Meta has shipped three different AI models that all share the same first name: Muse. Muse Glimmer, released August 10, 2026, is a small open-weight model built to run on your own GPU. Muse Spark 1.3, released September 2, 2026, is Meta's fast, cheap, cloud-hosted flagship for everyday chat and coding. Muse Wonder, released October 12, 2026, is a bigger, pricier, more agentic model Meta is calling its first genuinely frontier-class release. None of them replace each other, and the naming alone doesn't tell you which one actually fits what you're trying to do - so here's the real, practical difference.
Quick summary: Muse Glimmer (Aug 10, 2026) is a ~30-billion-parameter open-weight model, free to self-host, built to run agentic tasks offline on a single consumer GPU (24GB VRAM in 4-bit form). Muse Spark 1.3 (Sept 2, 2026) is Meta's lightweight, low-latency cloud flagship, priced at $1.25/$4.25 per million input/output tokens, scoring 61 on the Artificial Analysis Intelligence Index. Muse Wonder and Muse Wonder Full (Oct 12, 2026) are Meta's new, larger frontier tier, priced at $4/$20 per million tokens via API or $50-$200/month through a Muse Code subscription, built around persistent cross-session memory and long multi-step task persistence. All three remain available; Meta is not retiring any of them in favor of another.
The one-sentence version of each
- Muse Glimmer: the one you run yourself, for free, on your own computer, when privacy and offline use matter more than top-tier capability.
- Muse Spark 1.3: the one for everyday cloud chat and coding, fast and cheap, and Meta's most capable model for most ordinary use until Wonder arrived.
- Muse Wonder / Wonder Full: the expensive, memory-aware, long-task-persistent one, built for expert-domain work (law, finance, health) and complex agentic jobs, not casual chat.
| Model | Released | Access & Price | Built For |
|---|---|---|---|
| Muse Glimmer | Aug 10, 2026 | Free, self-hosted (Apache 2.0, ~30B params) | Local, offline, private agentic tasks on a consumer GPU |
| Muse Spark 1.3 | Sept 2, 2026 | $1.25 in / $4.25 out per 1M tokens (API/Muse Code) | Fast, cheap everyday chat and coding |
| Muse Wonder | Oct 12, 2026 | $4 in / $20 out per 1M tokens (API) | Expert-domain reasoning, frontier coding |
| Muse Wonder Full | Oct 12, 2026 | $50/mo (Power) or $200/mo (Maximum), Muse Code only | Full agentic capability with persistent cross-session memory |
The three models, one at a time
Muse Glimmer: the private, local option
Glimmer is the odd one out in the lineup, and deliberately so: it's a roughly 30-billion-parameter model distilled from Meta's larger, closed teacher model (originally Muse Spark), released under the permissive Apache 2.0 license so anyone can download, modify, and self-host it. In 4-bit quantized form, it fits inside 24GB of VRAM - the amount on a high-end consumer GPU like an RTX 4090 - and it runs fully offline, meaning your prompts, files, and screenshots never leave your machine. It was built specifically for agentic work: tool use, coding, reading files and screenshots, multi-step reasoning. The tradeoff is capability - Glimmer's ceiling sits below both Spark and Wonder, and Meta's own published benchmarks for it haven't been independently verified. For a deeper look at what it can and can't do, see our full review of Meta's new open Muse Glimmer.
Muse Spark 1.3: the everyday flagship
Spark 1.3, released a few weeks before Wonder, is Meta's cloud-hosted, proprietary model built for low latency and everyday usability rather than maximum raw capability. Meta's chief AI officer Alexandr Wang described it as Meta's biggest jump so far on coding and agentic work when it launched, claiming it was competitive with Claude Fable 5.1 and better than GPT-5.6 Sol specifically at software development - claims that are Meta's own but are at least anchored to a real independent number elsewhere: Spark 1.3 (in its publicly available xhigh setting) scores 61 on the Artificial Analysis Intelligence Index, a third-party benchmark, tying GPT-5.6 Sol (max) and Grok 4.6 (high) on that specific measure. Versus its own predecessor, Spark 1.2, Meta says 1.3 uses roughly 20% fewer tool calls and 25% fewer tokens to do comparable work - a real efficiency gain, not just a benchmark bump. For the fuller picture on how Meta's consumer-facing Muse product stacks up against ChatGPT, Claude, and Gemini, see our review, Meta Muse Spark: The Free AI Ranked 4th.
Muse Wonder: the new, expensive frontier tier
Wonder is where Meta is explicitly trying to compete at the very top of the market rather than on price. It's built around two features - Active Memory, which carries context about you across sessions and resurfaces relevant details later, and Active Focus, which keeps the model working through a hard multi-step task, retrying failed approaches and checking in before consequential actions - aimed squarely at long-horizon agentic work in expert domains like law, finance, and health, plus frontier-level coding. Meta claims Wonder beats Fable, Opus, and Astra on its own internal memory benchmark, though that specific number hasn't been published or independently verified. What is verifiable is the price: $4 input / $20 output per million tokens via API for Wonder, with Wonder Full - the fuller-capability variant - available only through a Muse Code subscription at $50 or $200 a month, not via API at all.
Don't assume "newest" means "the one you should use." Wonder is built and priced for a genuinely different job than Spark 1.3 - if your use case is everyday chat, quick coding help, or simple questions, Spark remains the cheaper, faster, and arguably more sensible choice; Wonder's premium is only worth paying if you actually need persistent memory across sessions or sustained multi-step agentic work.
A quick decision guide
- Want to self-host, run offline, and keep everything private, and have a decent consumer GPU? → Muse Glimmer.
- Want fast, cheap, capable everyday chat and coding through the cloud? → Muse Spark 1.3.
- Building a professional tool in law, finance, or health that needs to remember context across sessions and handle complex multi-step tasks? → Muse Wonder or Wonder Full, depending on whether API access or a flat-rate subscription fits your usage better.
- Not sure yet? Start with Spark 1.3 or Glimmer - both are dramatically cheaper ways to find out whether Meta's models fit your workflow before committing to Wonder's pricing.
01Is Muse Wonder better than Muse Spark 1.3?
It depends on the job. Wonder is larger, pricier, and built for long multi-step agentic work with persistent memory across sessions; Spark 1.3 is faster, cheaper, and built for everyday chat and coding. Meta is keeping both available rather than replacing one with the other.
02Which Muse model is actually free?
Only Muse Glimmer is free to use without limits - it's open-weight (Apache 2.0) and self-hosted, so there's no subscription or per-token fee, just your own hardware. Muse Spark 1.3 and Muse Wonder are both paid, cloud-hosted products.
03Can I run Muse Wonder on my own computer like Glimmer?
No. Wonder and Wonder Full are proprietary, cloud-hosted models available only through Meta's API or a Muse Code subscription - unlike Glimmer, there are no open weights to self-host.
04What's the cheapest way to try Meta's AI models?
Muse Glimmer, if you have a capable GPU (24GB VRAM in 4-bit form), since it's free to self-host. For a cloud option without a subscription commitment, Muse Spark 1.3's pay-as-you-go API pricing ($1.25/$4.25 per million tokens) is far cheaper than Wonder's $4/$20 rate or Wonder Full's $50-$200/month subscription.
Three overlapping model names from one company is a lot to track - and Meta isn't the only lab doing it; Anthropic, OpenAI, and Google all run similarly tiered lineups now. LumiChats cuts through that by giving you one subscription to chat with 40+ leading AI models side by side, including Meta's own lineup alongside Claude, GPT, and Gemini, so you can compare them on your actual questions instead of parsing a confusing tier chart.
