October 20 correction: Meta released Muse Spark Open on October 15: 816B total/42B active parameters, multimodal input, about 1M tokens of context and downloadable weights. The Muse Model License includes conditions, notably a separate agreement for qualifying Model-as-a-Service providers above $20 million aggregate revenue; it is not unrestricted MIT/Apache licensing. This invalidates the earlier prediction that Meta would release no more frontier open weights. The launch discussion below is historical, not a current model ranking.
Meta launched the original hosted Muse Spark on April 8. At launch it was natively multimodal, supported multi-agent “Contemplating” mode, and was built by Alexandr Wang’s team at Meta Superintelligence Labs.
That original hosted launch was closed, with no downloadable weights and invitation-only API access. Muse Spark Open is a separate release; downloading its weights does not imply unchanged hosted API access or pricing.
What changed
Llama was Meta’s open-source AI strategy for three years. The model family hit 1.2 billion downloads by early 2026. Developers built entire ecosystems on it — fine-tuning pipelines, local inference stacks, commercial products. Llama 4 Scout and Maverick shipped with MIT-equivalent licenses.
Muse Spark reverses all of that. The weights are proprietary. Meta says they “hope to open-source future versions” but provides no timeline. Gartner analyst Arun Chandrasekaran called it “a major shift” signaling Meta’s intention to move away from the Llama brand entirely.
Meta stock rose 9% on launch day — the strongest single-day response to a Meta product announcement in over two years. Wall Street clearly prefers the proprietary model.
Why it matters for developers
If you’re running Llama models locally or building products on Llama weights, nothing changes immediately. Llama 4 is still available, still MIT-licensed, still works.
But the signal is clear:
-
The closed-only prediction did not hold. Muse Spark Open provides a self-hosting route, subject to its own license and substantial hardware requirements.
-
The Llama ecosystem may stagnate. Without new open releases to drive community fine-tuning and tooling, the Llama ecosystem will gradually fall behind as other models improve.
-
Chinese open-source fills the gap. DeepSeek V4 (MIT, 1.6T params), Qwen 3.6 (Apache 2.0), MiMo V2.5 Pro (MIT, 1.02T params), and Kimi K2.6 are all open-weight and actively improving. The open-source AI leadership is shifting from Silicon Valley to China.
-
Self-hosting still needs a fit check. Compare license, full model memory and runtime support alongside alternatives such as Devstral and Granite. A weight release is not proof it fits your workstation.
What Muse Spark actually does
- Natively multimodal (text, image, voice)
- “Contemplating” mode: runs multiple reasoning agents in parallel before responding
- Health reasoning: trained with 1,000+ physicians on medical data
- Scores 52 on AI Intelligence Index (below GPT-5.4’s 57, above DeepSeek V4 Pro’s 52)
- Parameter count and architecture undisclosed
For coding specifically, Meta acknowledges “current performance gaps in long-horizon agentic systems and coding workflows.” It’s not a coding model — it’s a general assistant competing with ChatGPT and Gemini.
The original launch-era alternatives
| Provider | Best Open Model | License | Params |
|---|---|---|---|
| DeepSeek | V4 Pro | MIT | 1.6T (49B active) |
| Xiaomi | MiMo V2.5 Pro | MIT | 1.02T (42B active) |
| Alibaba | Qwen 3.6 | Apache 2.0 | Multiple sizes |
| Moonshot | Kimi K2.6 | Open | 1T+ |
| Mistral | Medium 3.5 | Apache 2.0 | 128B |
| IBM | Granite 4.1 | Apache 2.0 | 3B/8B/30B |
| Meta | Llama 4 (legacy) | MIT-equiv | 405B |
Those alternatives remain relevant, but Muse Spark Open means this is no longer a landscape without Meta’s downloadable weights.
What to do now
- If you’re on Llama 4: Evaluate any migration separately. Muse Spark Open is not a promise about a future Llama release.
- If you need frontier open-source: Look at DeepSeek V4 Pro or MiMo V2.5 Pro. Both are MIT-licensed and competitive with closed models.
- If you’re building local AI infrastructure: Diversify away from any single provider. The Chinese model ecosystem is your best bet for continued open releases.
- If you’re using Meta AI (the product): Muse Spark will roll out across WhatsApp, Instagram, Facebook, and Messenger. It’ll be free to use through Meta’s apps — just not self-hostable.
The practical decision is now which license, model and hardware fit your workload, not whether Meta has abandoned all downloadable weights.