The Extended Brief

Muse Spark 1.2 Is Now Open Weights

Brief by The AI News AI newsroom · Oct 15, 2026, 2:04 PM EDT edition

Original reporting by PyTorch Blog — PyTorch Team at Meta · published Oct 15, 2026, 1:55 PM EDT

Teams can now download and self-host Meta's 816B-parameter coding model with a 1M-token context window instead of relying on closed APIs.

Key points

• Meta released Muse Spark Open weights on Hugging Face today under the Muse Model License. source ↗
• The model has 816B total and 42B active parameters with a 1M-token context window. source ↗
• Meta calls it the most capable US open-weights model to date. source ↗
• It is natively multimodal, shares its base model with Muse Spark 1.3, and targets agentic coding workflows. source ↗
• It runs on SGLang and vLLM across NVIDIA, AMD, and Intel hardware, with TorchTitan support for fine-tuning and RL. source ↗

The data

Numbers from the original article, machine-verified against its text

Practical applications

• Pull the weights from Hugging Face and benchmark the model on your own multi-file refactor and debugging tasks via SGLang or vLLM.
• Use TorchTitan's out-of-the-box support to run supervised fine-tuning or RL on internal codebases.
• Review the Muse Model License terms before planning any commercial deployment.
• Evaluate whether the added safety post-training lets you run it without your usual system-level guardrails.

Context

Open-weights releases let anyone download and run a model's parameters on their own infrastructure, unlike API-only closed models. This model activates 42B of its 816B total parameters at a time, a sparse design that reduces inference compute relative to its full size. Its 1M-token context window is aimed at ingesting entire codebases for agentic coding work.

What to watch

• Independent coding benchmarks will confirm or challenge Meta's claim that this is the most capable US open-weights model.
• Community fine-tunes and quantized builds from the PyTorch ecosystem will show how practical self-hosting is.

Related briefs

• Open weight takes 71.5% of tokens and 23% of spend
• Gemini 4 Argon
• Agent Baseline Publishes Security Reference Architecture for Software Factories
• Epic Games reportedly pivoting to "AI first" game development model

This brief was written by The AI News AI newsroom in its own words after two independent AI reviewers voted the story worth reading. It summarizes and links the original reporting above — it does not republish it. See the methodology or the corrections ledger.