Breaking02 / IT
Meta launches Muse Spark Open, its 816B agentic coding model
The model lists 42B active parameters and a 1M-token context window, but the introduction to NVIDIA's blog post gives no licence terms or local running requirements.

Meta launched Muse Spark Open, an 816B-parameter model built for coding, NVIDIA said in a developer blog post on 16 Oct 2026. The model targets developers who build agentic coding workflows, in which AI systems carry out multi-step programming tasks.
The blog post describes a mixture-of-experts (MoE) architecture, a design that uses only part of a model's parameters for each input. It lists 42B active parameters. The context window, the amount of input a model can handle at once, is given as 1M tokens.
According to NVIDIA, the model was trained on NVIDIA GB300 NVL72 Blackwell Ultra. Training was aimed at agentic coding workflows.
NVIDIA said the model delivers higher first-attempt accuracy and reliable tool calling, the ability to invoke external software functions. It also named understanding of codebases across multiple repositories as a strength.
Licence and local use. The introduction to NVIDIA's post does not state the baseline for the first-attempt accuracy claim, the licence terms or the specifications for running the model locally.