Muse Spark 1.3
Meta (Meta Superintelligence Labs) · Multimodal model
Muse Spark is the proprietary model line Meta introduced in April 2026 to replace Llama inside Meta AI; version 1.3 arrived September 2, 2026. It is aimed at coding and agents, has a 1M-token context, and is sold cheaply through…
Current
ReleasedSep 2, 2026
Context window1M tokens
Input → outputtext, image, video, pdf → text
API pricemuse-spark-1.3: $1.25 / 1M input, $0.15 / 1M cached input, $4.25 / 1M output (not used to improve Meta's products). muse-spark-1.3-contributor: $0.10 input, $0.002 cached input, $0.20 output per 1M tokens (used to improve Meta's products).
Consumer accessMuse Spark (original 1.x line, from April 2026) powers Meta AI in the Meta AI app and website; the 1.3 page itself does not mention consumer access.
Best for
Cost-efficient long-horizon coding and agentic work with very long context.
Overview
Muse Spark is the proprietary model line Meta introduced in April 2026 to replace Llama inside Meta AI; version 1.3 arrived September 2, 2026. It is aimed at coding and agents, has a 1M-token context, and is sold cheaply through the Meta Model API. It now sits below the newer Muse Wonder.
Strengths and limits
- Large long-context gains over 1.2 (MRCR 512K–1M: 98.1)
- Cheap API, with a very low-cost 'contributor' tier
- Meta reports ~25% fewer tokens on coding tasks vs 1.2
- Closed weights; contributor tier lets Meta use your data
- Meta notes audio understanding in 1.3 is not fully supported (recommends 1.2 for audio)
Published benchmark scores
DeepSWE v1.1 (max)75.4
Source: dev.meta.ai
Terminal-Bench 2.1 (max)88.8
Source: dev.meta.ai
OSWorld 2.0 partial / binary (max)66.9 / 32.0
Source: dev.meta.ai
MRCR 512K–1M (max)98.1
Source: dev.meta.ai
Scores are as published by the source; settings (reasoning effort, tools, attempts) differ between labs, so compare with care. How to read benchmarks →