$ cat wiki/models/muse-spark.md
Muse Spark (1.0 / 1.1)
modelupdated 2026-07-21created 2026-05-16
Spec
| Attribute | Value |
|---|---|
| Developer | Meta / Meta Superintelligence Labs (MSL) |
| Released | 1.0: 2026-04-08 · 1.1: 2026-07-09 |
| Announced | 1.0: 2026-04-08 · 1.1: 2026-07-09 |
| Context window | 1,000,000 tokens (1.1) |
| Pricing | $1.25/M input · $4.25/M output (Meta Model API, public preview) |
| License | proprietary (closed / API-only, not open-weights) |
| Availability | Meta Model API (public preview, US-based developers only) |
| Positioning | "Personal Superintelligence" (1.0); agentic coding + professional domains (1.1) |
Release Date
- 1.0: 2026-04-08
- 1.1: 2026-07-09
Muse Spark 1.1 (July 2026)
Major upgrade released alongside the Meta Model API — Meta's first paid external AI product.
Benchmarks (1.1)
- MedScribe: new SOTA (beats Claude Fable 5; 10× cheaper, 2× faster)
- TaxEval: new SOTA
- Harvey's Legal Agent Bench: new SOTA
Key Improvements (1.0 → 1.1)
- Major gains in tool use, computer use, and coding
- Zero-shot generalization to new native tools, MCP servers, and custom skills
- Enhanced multimodal understanding
- Cybersecurity-aware deployment mitigations
Meta Model API
- Price: $1.25/M input, $4.25/M output
- SDK compatibility: OpenAI Chat Completions/Responses format AND Anthropic Messages format
- $20 free credits for new US developers
- First paid external AI product from Meta
Muse Spark 1.0 (April 2026)
Meta's official framing is "Scaling Towards Personal Superintelligence".
Announced the same day: "Scaling how we build and test advanced AI" — disclosing infrastructure and methodology. Closed / API-only (not open-weights), distinguishing it from Llama.
Significance
- MSL's first external signal from Meta (1.0)
- 1.1 marks Meta entering the commercial frontier AI API market for the first time
- "Personal superintelligence" is a term Meta has used to differentiate from LeCun's world model stance
- A signal of Meta AI's LLM-centric shift under MSL (led by Alexandr Wang)
Open Questions
- Does Muse Spark 1.1 use Watermelon as its backbone, or is it a separate model line?
- What is the relationship to Llama 4/5 and open-weights strategy?
- Will Meta Model API expand outside the US?
Compared To
| Model | Price (in/out per Mtok) | Context | Key Bench |
|---|---|---|---|
| Muse Spark 1.1 | $1.25/$4.25 | 1M | MedScribe SOTA, TaxEval SOTA |
| Claude Sonnet 5 | $2/$10 (promo) | 1M | SWE-bench Pro 63.2% |
| GPT-5.6 Terra | $2.50/$15 | 4M | — |
| Grok 4.5 | $2/$6 | 2M | SWE-Pro 64.7% |
- Claude Opus 4.7 (Anthropic April announcement) — same cluster
- GPT-Rosalind (OpenAI April announcement) — same week