Meta has finally introduced the model the industry has been waiting nine months to see. Meta Muse Spark — the first flagship release from Meta Superintelligence Labs under chief AI officer Alexandr Wang — landed in early May 2026 with a 262K-token context window, native multimodal input across voice, text, and images, and benchmark numbers strong enough to put Meta back inside the conversation it was visibly losing through 2025. It is also the first time Meta has shipped a frontier model behind a closed weights policy. That single decision matters as much as anything in the spec sheet.
What’s actually new
The substance of the launch is in the architecture, the benchmarks, and the strategic break from Llama precedent.
- Closed weights, not open. Every previous Meta frontier model — Llama 1 through Llama 4 — shipped with downloadable weights. Muse Spark does not. It is API-only, with selective enterprise hosting partnerships. The Llama brand survives for smaller open-weight releases; the frontier flagship has moved to the closed lane.
- Natively multimodal input. Voice, text, and images go straight into the model in a single token stream. There is no separate vision encoder bolted on — the model trains end-to-end across modalities. Output is text plus structured tool calls; no image or audio generation in this release.
- 262K context window. A meaningful step beyond the 128K and 200K windows that have been industry standard since 2024. Long-document analysis, multi-document RAG, and full-codebase reasoning are all directly addressable without chunking.
- Top-five frontier benchmark slot. Muse Spark ranks fourth on Artificial Analysis Intelligence Index v4.0 with a score of 52, sitting behind only GPT-5.4 (57), Gemini 3.1 Pro (57), and Claude Opus 4.6 (53). It is the first Meta frontier model to land in the leading cohort.
- HealthBench Hard leadership. On the medical-reasoning HealthBench Hard suite, Muse Spark scores 42.8 — ahead of GPT-5.4’s 40.1 and Gemini 3.1 Pro’s 20.6. Meta is signaling that health and life-sciences applications are an opening pricing wedge.
- Built under the new org. This is the first model fully shipped by Meta Superintelligence Labs since Alexandr Wang took over and Meta took a 49% stake in Scale AI for $14.3 billion. The organizational reset is producing visible output.
Why it matters
- Meta is back in the frontier race. Through most of 2025 the consensus was that Meta had been left behind by GPT, Gemini, and Claude. Muse Spark does not unseat the leaders, but it makes the race four-way again. Decisions about which model to integrate now have a real Meta option for the first time in 18 months.
- The closed-weights pivot reshapes the open-source AI map. Meta was the largest source of open frontier weights. With the flagship moving closed, the practical leaders in open-weights frontier capability shift toward Mistral, DeepSeek, Qwen, and the Llama lineage’s lower tiers. Companies whose AI strategy assumed Meta would keep shipping weights need a new plan.
- $115-135 billion of capex sits behind it. Meta disclosed AI capital expenditure guidance of $115-135 billion for 2026 alongside the Muse Spark launch — nearly double 2025. The model is the visible deliverable on the largest single corporate AI infrastructure bet of the year.
- Health is a competitive opening. The HealthBench Hard delta is striking. Meta has positioned Muse Spark as health-capable in a way that GPT and Gemini are not, which suggests deliberate investment in medical reasoning data and reinforcement signals. Healthcare AI buyers should evaluate.
- Native multimodality changes the integration surface. Apps that today route voice through Whisper, images through CLIP, and text through a chat model can collapse three calls into one with Muse Spark. That simplification is real engineering value beyond raw quality.
- Wang’s playbook is visible. The closed model, the data-centric training story, the focus on benchmarks that competitors do not own, and the simultaneous capex disclosure all reflect Alexandr Wang’s data-and-evals-first approach from his Scale AI tenure. This is not a refresh; it is a strategy change.
How to use it today
Muse Spark is in private preview through Meta’s API as of mid-May 2026, with broader access rolling out through enterprise partnerships and a waitlist for individual developers. The integration patterns to know:
- Request preview access. Apply through the Meta AI for Developers portal. Approval is currently weighted toward enterprises with existing Meta business relationships, but solo developers with credible use cases are getting in within 2-3 weeks.
- Authenticate against the new endpoint. The Muse Spark API is OpenAI-compatible at the schema level, which means most existing client libraries work with a base-URL swap.
import os from openai import OpenAI client = OpenAI( api_key=os.environ["META_AI_API_KEY"], base_url="https://api.meta.ai/v1", ) response = client.chat.completions.create( model="muse-spark", messages=[ {"role": "system", "content": "You are a clinical decision support assistant."}, {"role": "user", "content": "Patient: 58F, BMI 31, A1c 7.4. Outline a 12-week treatment plan."} ], temperature=0.3, max_tokens=2048, ) print(response.choices[0].message.content) - Send multimodal input in one request. Audio, image, and text can be combined in a single message; do not split.
response = client.chat.completions.create( model="muse-spark", messages=[{ "role": "user", "content": [ {"type": "text", "text": "What is shown in this chest X-ray and audio note?"}, {"type": "image_url", "image_url": {"url": "https://example.com/cxr.png"}}, {"type": "input_audio", "input_audio": { "data": "", "format": "wav" }} ] }] ) - Use the long-context window deliberately. 262K tokens is roughly 200,000 English words — a full code repository or a 600-page document. Muse Spark performs best when prompts include relevant retrieved chunks rather than the entire haystack; do not abandon RAG, but you can be lazier about it.
- Stream for latency-sensitive UX. Time-to-first-token on Muse Spark is in the 600-900 ms range for typical prompts. Streaming responses to the user keeps perceived latency low even when the full completion runs 8-15 seconds.
stream = client.chat.completions.create( model="muse-spark", messages=messages, stream=True, ) for chunk in stream: if chunk.choices[0].delta.content: print(chunk.choices[0].delta.content, end="", flush=True) - Run a parallel A/B against your incumbent. Whatever frontier model you run today, set up a 1-3% traffic mirror to Muse Spark and grade outputs offline. Pricing is competitive but not the cheapest in the field; the decision is quality-per-dollar on your specific workload, not list price.
How it compares
The Artificial Analysis Intelligence Index v4.0 is the most-watched cross-lab benchmark suite for general capability. The post-launch standings:
| Model | AAII v4.0 score | Context window | Native modalities (input) | Open weights |
|---|---|---|---|---|
| GPT-5.4 | 57 | 1M tokens | Text, image, audio | No |
| Gemini 3.1 Pro | 57 | 2M tokens | Text, image, audio, video | No |
| Claude Opus 4.6 | 53 | 500K tokens | Text, image | No |
| Meta Muse Spark | 52 | 262K tokens | Text, image, audio | No |
| DeepSeek-V4 | 49 | 128K tokens | Text, image | Yes |
| Qwen 3.5 | 47 | 256K tokens | Text, image | Yes |
The HealthBench Hard scores tell a more interesting story for verticals where Muse Spark may be the right pick:
| Model | HealthBench Hard | Notes |
|---|---|---|
| Meta Muse Spark | 42.8 | Best-in-class on this suite |
| GPT-5.4 | 40.1 | Close second |
| Claude Opus 4.6 | 34.7 | Mid-pack |
| Gemini 3.1 Pro | 20.6 | Notably weaker on medical reasoning |
The takeaway: for general agentic and reasoning tasks, GPT-5.4 and Gemini 3.1 Pro remain the strongest single-model choices. For health-domain applications, Meta Muse Spark is genuinely competitive at the top of the field. For latency-sensitive multimodal ingestion, the native voice+image+text input pipeline is a real architectural advantage that flattens integration complexity.
What’s next
Three near-term threads to watch.
Pricing pressure on the frontier tier. Meta has not published list pricing in detail, but signaling suggests aggressive enterprise pricing to win share fast. If Muse Spark lands at meaningfully below GPT-5.4 and Claude Opus 4.6 on a per-million-token basis, the entire frontier price floor moves down. Procurement teams should renegotiate frontier contracts in Q3 2026 with the new comparable in hand.
The Muse series cadence. The label “Muse Spark” implies more Muse models coming. Industry consensus expects a Muse Forge (larger, training-focused) and Muse Lattice (smaller, on-device) within 12 months, mirroring the OpenAI o-series and Anthropic Claude tier structure. Plan for a Meta product family, not a single SKU.
Open-weights fallout. Meta’s frontier closure puts pressure on the open-weights ecosystem to consolidate. Mistral and DeepSeek are likely beneficiaries. Llama 4 will continue to receive maintenance releases at the smaller tiers, but the frontier-quality open weights baton has moved out of Meta’s hands. Teams whose deployment relies on running frontier weights privately need to evaluate their alternatives now.
Frequently Asked Questions
Is Meta Muse Spark open-weights like Llama?
No. Muse Spark is closed-weights and API-only. It is the first frontier-tier Meta model to ship without downloadable weights. The Llama brand continues for smaller open-weights releases, but the frontier flagship has moved to the closed model.
How does Muse Spark compare to GPT-5.4 and Claude Opus 4.6?
On the Artificial Analysis Intelligence Index v4.0, Muse Spark scores 52 versus GPT-5.4’s 57 and Claude Opus 4.6’s 53 — a small but real gap on general capability. On HealthBench Hard for medical reasoning, Muse Spark leads at 42.8 versus GPT-5.4’s 40.1 and Claude’s 34.7. Choose Muse Spark for health applications and multimodal ingestion; choose GPT-5.4 or Claude for the broadest general agentic tasks.
What is the context window for Meta Muse Spark?
262,144 tokens (262K). That is roughly 200,000 English words, a full mid-sized code repository, or a 500-700 page document. The window is shorter than GPT-5.4’s 1M or Gemini 3.1 Pro’s 2M but longer than Claude Opus 4.6’s 500K.
Can Muse Spark process audio and images?
On the input side, yes — voice, images, and text can all go into a single API call in one tokenized stream. On the output side, Muse Spark produces text and structured tool calls only; this release does not generate images, audio, or video.
How do I get access to the Muse Spark API?
Apply through Meta AI for Developers. As of mid-May 2026, access is gated by waitlist with priority for enterprise customers and existing Meta business partners. Solo developers with strong use cases are reporting 2-3 week approval timelines. Public general availability is expected in Q3 2026.
Why did Meta drop the open-weights playbook for Muse Spark?
Meta has not stated a single reason publicly, but three factors line up. The cost of training frontier-class models has risen sharply; recouping it through API revenue is harder when weights are public. Alexandr Wang’s prior strategy at Scale AI was data-and-evals-first, which fits a closed model better. And the competitive risk of giving frontier weights to overseas labs has been an explicit policy concern voiced by Meta executives in 2026.