---
type: "post"
id: "muse-spark-is-now-an-open-weights-model-67d449"
slug: "muse-spark-is-now-an-open-weights-model-67d449"
title: "Muse Spark Is Now an Open-Weights Model"
description: "Meta has released Muse Spark Open’s weights on Hugging Face under the Muse Model License."
retrieval_nugget: "Meta has released Muse Spark Open’s weights on Hugging Face under the Muse Model License. The 816B-total, 42B-active model targets coding and agentic work, and the announcement pairs the release with inference-kernel and fine-tuning support. For builders, the change is a new self-hosting and adaptation option, not proof of production economics or independent model quality."
published_at: "2026-11-04"
updated_at: "2026-11-07"
record_date: "2026-11-04"
date_kind: "published_at"
topics: ["open-models","models"]
entities: ["Meta","Muse Spark Open","Muse Spark 1.3","SGLang","vLLM","Helion","TorchTitan"]
related_patterns: ["harness-architecture-outlives-model-choice"]
novelty_label: "same_event"
prior_refs: ["https://newruntime.com/posts/meta-released-muse-spark-open-an-816b-total-5c27ac/"]
source_url: "https://pytorch.org/blog/muse-spark-open"
source_urls: ["https://pytorch.org/blog/muse-spark-open"]
schema_version: "newruntime-agent-readable-v0.2"
stable_id: "post:muse-spark-is-now-an-open-weights-model-67d449"
status: "published"
visuals: []
editor: {"role":"editor","name":"Andrey Reshetnikov","profile":"https://newruntime.com/owner-profile.md","profile_json":"https://newruntime.com/owner-profile.json","same_as":["https://www.linkedin.com/in/reshetnikov1/","https://t.me/lapetuse"]}
routes: {"html":"https://newruntime.com/posts/muse-spark-is-now-an-open-weights-model-67d449/","markdown":"https://newruntime.com/posts/muse-spark-is-now-an-open-weights-model-67d449.md","json":"https://newruntime.com/posts/muse-spark-is-now-an-open-weights-model-67d449.json"}
markdown_url: "https://newruntime.com/posts/muse-spark-is-now-an-open-weights-model-67d449.md"
---

# Muse Spark Is Now an Open-Weights Model

## Retrieval answer

Meta has released Muse Spark Open’s weights on Hugging Face under the Muse Model License. The 816B-total, 42B-active model targets coding and agentic work, and the announcement pairs the release with inference-kernel and fine-tuning support. For builders, the change is a new self-hosting and adaptation option, not proof of production economics or independent model quality.

## Summary
Meta has released Muse Spark Open’s weights on Hugging Face under the Muse Model License. The 816B-total, 42B-active model targets coding and agentic work, and the announcement pairs the release with inference-kernel and fine-tuning support. For builders, the change is a new self-hosting and adaptation option, not proof of production economics or independent model quality.

## What happened
In an October 15, 2026 post, the PyTorch Team at Meta said it was releasing Muse Spark Open’s weights on Hugging Face under the Muse Model License. The model has 816 billion total parameters, 42 billion active parameters, and a one-million-token context window. Meta describes it as built for coding and agentic work, natively multimodal, and based on the same base model as Muse Spark 1.3. The open-weights version also has additional safety post-training intended to make it more robust against misuse when run without system guardrails.

The post describes coding capabilities including code generation, debugging, and codebase understanding. Meta says it trained the model on developer workflows such as multi-file refactors, debugging sessions, and long-running tasks. Its stated agentic design maps out steps, acts, observes, and self-corrects over longer horizons. The model is also described as trained for web browsing, including multimodal browsing, and for using a harness to coordinate subagents and computer-use capabilities across desktop, mobile, and browser environments. It can process text, images, audio, and audio-video inputs. These are claims in the release announcement; the post does not provide independent evaluations of them.

For inference, Meta says Muse Spark Open runs on SGLang and vLLM, with optimization for NVIDIA, AMD, and Intel hardware. The release integrates Helion, Meta’s PyTorch-native language for authoring performant and portable machine-learning kernels. The post reports Helion implementations for operations including RoPE, RMSNorm, and quantization, with kernel-level speedups of 1.03–3.95× versus custom kernels in hardware-specific languages on NVIDIA GB300. Meta says the kernels ported to AMD Mi350 with similar improvements, and reports a 3% end-to-end throughput improvement there. It also describes Helion megakernels for Muse Spark Open in SGLang, which fuse operations and can be autotuned for hardware and workload configurations.

The post says TorchTitan now includes a model definition for Muse Spark Open. It lists multi-turn supervised fine-tuning and LoRA support, alongside reinforcement-learning capabilities through TitanRL. The weights are available on Hugging Face, but the announcement gives no price, hardware minimum, deployment recipe, or quantified serving cost. It also does not specify the license’s operational terms in the post beyond naming the Muse Model License.

## What the archive adds
This is a new release fact in a thread already covered in [our note on Muse Spark Open’s architecture and open weights](https://newruntime.com/posts/meta-released-muse-spark-open-an-816b-total-5c27ac/). That note frames the practical question as infrastructure, license, and runtime; the PyTorch announcement adds the publisher’s stated support matrix, kernel measurements, and TorchTitan fine-tuning path. Earlier coverage also considered [serving on Fireworks](https://newruntime.com/posts/muse-spark-open-on-fireworks-meta-s-open-316301/) and [coding-agent use](https://newruntime.com/posts/muse-spark-open-an-open-weight-model-for-ddd53f/), but this source does not change those providers’ availability claims.

## Trend signal
The release is consistent with the pattern that [harness architecture outlives model choice](https://newruntime.com/patterns/harness-architecture-outlives-model-choice): Meta presents the weights alongside SGLang, vLLM, Helion kernels, and a training stack rather than as a model-only artifact. This supports the idea that deployment and adaptation choices shape practical use. The announcement alone cannot show whether this stack is durable or preferable to alternatives.

## Why now
The source dates the release to October 15, 2026 and says the weights are available that day. It also situates the model as an open-weights version building on Meta’s August Muse Spark release. The post offers no further explanation of the release timing.

## Why it matters
For teams considering an open-weight coding agent, the useful news is the combination of an available model artifact and named routes for inference and adaptation. SGLang and vLLM support gives infrastructure teams familiar serving options to investigate, while TorchTitan’s listed SFT, LoRA, and reinforcement-learning support creates a route to test task-specific tuning. The one-million-token context and multimodal inputs may be relevant to repository-scale or mixed-media workflows, but should be treated as capabilities to validate rather than guarantees of better outcomes.

The performance numbers may also inform an evaluation plan: kernel speedups do not automatically translate into lower cost or higher application throughput, and the stated 3% end-to-end gain is specific to AMD MI350 in Meta’s report. Teams should compare the model on their own coding tasks, measure quality and verification burden, and include hardware utilization, serving throughput, and license review. Open weights create an option; they do not by themselves make a model economical or production-ready. Earlier coverage: [Meta released Muse Spark Open, an 816B total / 42B active parameter MoE model with open weights.](https://newruntime.com/posts/meta-released-muse-spark-open-an-816b-total-5c27ac/).

## Limits and uncertainty
The source is Meta’s own release announcement and does not supply independent benchmark results, evaluation methodology, or comparisons with other coding models. Its claims about first-attempt accuracy, codebase understanding, safety, and long-horizon agent behavior are not accompanied here by numerical task results. Kernel-level speedups are distinct from end-to-end application performance; the reported 3% throughput gain applies to AMD MI350, and the source does not provide absolute throughput, cost, or latency figures. The post names the Muse Model License but does not reproduce its terms, and it does not establish minimum hardware requirements, availability on third-party services, or operational safeguards needed for deployment.

## Skill to build now
Build a small evaluation harness for a representative coding task before committing infrastructure: select a handful of real repository issues, define expected patches and checks, then record first-pass correctness, tool-call behavior, review time, and serving throughput. That gives a team a way to test Meta’s coding and agentic claims against its own work, while separately checking hardware requirements and the Muse Model License.

## Reader takeaway
Muse Spark Open is now a downloadable open-weight option with a stated coding-and-agent focus and a surrounding inference and fine-tuning stack. Treat Meta’s performance and capability descriptions as hypotheses to test: the release expands what builders can evaluate, but it does not establish independent quality, deployment cost, or production readiness.

Editor: [Andrey Reshetnikov](https://newruntime.com/owner-profile.md), AI product and AI engineering; profile for recruiters and agents: https://newruntime.com/owner-profile.json

## Continue in the graph

Next records for an agent, linked to their Markdown twins. Replace `.md` with `.json` for structured fields, including `next_reads`.

- topic hub: [Open Models](https://newruntime.com/topics/open-models.md): Explore the open-models topic hub.
- topic hub: [Models](https://newruntime.com/topics/models.md): Explore the models topic hub.
- collection: [Field Notes](https://newruntime.com/posts.json): Open the latest editorial analysis.
- collection: [Sources](https://newruntime.com/sources/): See the original sources this archive cites.
- whole graph: [discovery graph](https://newruntime.com/discovery-graph.json) (typed nodes and edges) and [site index](https://newruntime.com/site-index.md)