Glonce

Meta ships Muse Spark 1.3 for agentic coding

modelsThe Glonce editorsHacker News🔥78Glonce rating78Impact60Novelty50Relevance100Surprise45Players90Credibility85

Meta has put up a product page for Muse Spark 1.3, the latest entry in its Muse Spark model line, listed alongside earlier Muse Spark 1.2 and 1.1 versions and sitting next to Muse Glimmer, Muse Image, and Muse Voice Transcribe as well as the Llama 4 and Llama 3 families in Meta's model API. Meta describes the model as trained for agentic workflows and optimized for competitive coding performance, with higher first-attempt accuracy and reliable tool calling. The page frames it specifically for long-horizon, agentic use: it tracks context and prior results, works through messy or conflicting inputs, and asks for input when needed. Meta also lists native multimodal perception for video, images, and documents, saying visual reasoning runs through a real execution environment.

The model is offered in two tiers, both with a 1M-token context window. The standard muse-spark-1.3 is priced at $1.25 per million input tokens, $0.15 per million cached input tokens, and $4.25 per million output tokens. A cheaper muse-spark-1.3-contributor variant costs $0.10 per million input tokens, $0.002 per million cached input tokens, and $0.20 per million output tokens.

The page carries no release date, no benchmark scores, no comparison to other models, and no named individuals behind the release. It also does not explain model size, architecture, or training data.

Key facts

  • Meta published a product page for Muse Spark 1.3, part of its Muse model lineup alongside Muse Glimmer, Muse Image, Muse Voice Transcribe, and the Llama 4 and Llama 3 families.
  • The model ships in two tiers, both with a 1M-token context window: the standard muse-spark-1.3 and the cheaper muse-spark-1.3-contributor.
  • Standard pricing is $1.25 per million input tokens, $0.15 per million cached input tokens, and $4.25 per million output tokens.
  • Contributor pricing is $0.10 per million input tokens, $0.002 per million cached input tokens, and $0.20 per million output tokens.
  • Meta describes the model as built for long-horizon agentic workflows with native multimodal perception for video, images, and documents, and says visual reasoning runs through a real execution environment.

Why it matters

Muse Spark 1.3 extends Meta's Muse Spark line, following the 1.2 and 1.1 releases listed alongside it on Meta's model page, and it sits next to Muse Glimmer, Muse Image, and Muse Voice Transcribe as well as the Llama 4 and Llama 3 families in Meta's model API. Meta positions it specifically for agentic and coding work rather than general chat, built around first-attempt accuracy and reliable tool calling, the two properties that matter most when a model is calling external tools instead of just answering questions.

Who it affects

Anyone building agent or coding tools on top of Meta's model API, particularly teams sensitive to per-token cost. The contributor tier's input price of $0.10 per million tokens sits far below the standard tier's $1.25, giving smaller projects or high-volume agent loops a cheaper entry point without switching model families.

How to use it

The model is offered through Meta's Model API in two variants. muse-spark-1.3 costs $1.25 per million input tokens, $0.15 per million cached input tokens, and $4.25 per million output tokens, with a 1M-token context window. muse-spark-1.3-contributor uses the same 1M-token window at $0.10 per million input tokens, $0.002 per million cached input tokens, and $0.20 per million output tokens. No release date or general-availability details appear on the page.

How solid is it

Everything here comes from Meta's own product page, which did not fetch cleanly on the first attempt (it returned an HTTP 400) before a follow-up fetch pulled the content used here. There is no independent benchmark, no named engineer or executive behind the release, and no comparison to competing models. Every capability claim (agentic performance, coding accuracy, tool calling, multimodal perception) is Meta's own description of the product, not a measured result.

Risks and caveats

The claims of higher first-attempt accuracy and reliable tool calling are vendor language with no accompanying benchmark or methodology, so they read as marketing positioning rather than a verified performance figure. The page gives no release date, no model size, and no architecture detail.

“trained for agentic workflows and optimized for competitive coding performance”

— Meta, Muse Spark 1.3 product page