Muse Spark Open
Released 15 October 2026 with downloadable weights — 816 billion parameters, 42 billion active, 1M-token context — under Meta’s Muse Model License, which Meta says has “limited commercial terms”. Meta calls it “The open-weight version of Muse Spark”, given “additional safety post-training” for use without system guardrails, and released it two months after the 30B Muse Glimmer and three days after Muse Wonder, following weeks of users asking when Muse Spark weights would appear.
Thin — one week old. The record is Meta’s release post, the model card (rendered on the gated Hugging Face page) and an evaluation-methodology note. No corpus matches yet (supplement db through 2026-10-21). Sourcing skew, named: entirely official so far.
Sources
Official
- 2026-10-15 Introducing Muse Spark Open, The Most Capable US Open‑Weight Model to Date (Meta) — “Today, we’re releasing the weights for Muse Spark Open, the most capable US open-weight model to date. Muse Spark Open is a natively multimodal model with 816B total parameters (42B active) and a 1M-token context window.” On running it locally: “Muse Spark Open is the first model in its class to run on your desktop using two Apple M5 Ultra Mac Studios with the latest MLX framework.” On safety: “Open weights also carry their own considerations, and Muse Spark Open has additional safety post-training for that reason.” On license: “We’re releasing the weights under the Muse Model License with limited commercial terms, which allow many developers to use the model for free.”
- 2026-10-15 Muse Spark Open model card (Hugging Face, gated download; card text rendered on the page) — “This open weights version has additional safety post-training to make it more robust against misuse as a standalone model without system guardrails.” Mixture-of-Experts transformer with a perception encoder; 256 experts, 8 active per token; input text, image, video and audio, output text. Collection: meta-models/muse-spark-open.
- 2026-10-15 Muse Spark Open Evaluation Methodology (PDF) — comparison settings against other open-weight models (including Inkling, Kimi K3, GLM 5.3 and Qwen3.8).
- 2026-10-02 Developing capable models responsibly (Meta) — the framework revision covering open releases: “Someone holding the weights can resample to get around a refusal, prefill outputs, or fine-tune refusal behavior away.”
- tk — the full evaluation report (promised “soon”); the Muse Model License text; independent benchmarks.
Writing & commentary
tk — no independent coverage retrieved in this pass.
Tweets
No corpus matches on muse spark open / muse-spark-open / spark open through 2026-10-21. The weights-release requests that preceded it are recorded on the Muse Spark page (@jmbollenbacher, 2026-09-06 and 10-12).
Official record
- Weights released 15 October 2026 on Hugging Face (gated), Amazon SageMaker AI, Google Cloud Model Garden and Microsoft Foundry Models, in BF16 and NVFP4; hosted inference through Baseten, Fireworks AI, Groq, Modal, Nebius, OpenRouter and Together AI. License: Muse Model License. CONFIRMED
- Model card: 816B total / 42B active parameters; 62 layers; 256 experts with 8 active per token; a 1.8B-parameter perception encoder; context “1M at inference”; vocabulary 202,048; “Multilingual, trained on data from more than 100 languages”; self-hosting needs “a GPU with at least 700GB of VRAM for NVFP4 and 2TB for BF16.” The published chat template defaults
cutoff_dateto2026-01-04. CONFIRMED - Headline numbers as published (Meta’s table, xhigh effort): Artificial Analysis Intelligence Index 39 (GLM 5.3 45, Kimi K3 44, Qwen3.8 40, Inkling 26); MCP-Atlas 89.2. CONFIRMED (as published)
- Jailbreak robustness (model card, Safety rows): on StrongREJECT v2 (adaptive), lower is better, Muse Spark Open 0.5 against Inkling 86.4, Nemotron 3 Ultra 80.7, Kimi K3 66.6, Qwen3.8 65.6, GLM 5.3 62.9. On cyber misuse with adaptive attackers and no system safeguards: “The model refused 99.2% of jailbreaks on our internal benchmark and 94.8% when aggregating across PolyGuard (cyber subset), RedCode, and CyberBench 2024.” CONFIRMED (as published)
- Loss of control (model card): “On our misalignment evaluations, the model showed low propensity for deceptive or unilateral harmful behavior, with a 98.1% honesty rate on the MASK benchmark and a 0% misbehavior rate on the Agentic Misalignment benchmark.” The card adds: “Downstream modifications could degrade safety, and results for the released checkpoint should not be assumed to apply to modified models.” CONFIRMED
- Deployment guidance (model card): “For user-facing applications, developers should avoid exposing raw chain-of-thought and apply appropriate safety checks to any reasoning summaries.” A noted quirk: with very short or low-resolution videos, “the model may sometimes reply that it can't view the video.” CONFIRMED
History
- 2026-04-08 → 09-02 Closed first: Muse Spark and its point releases ship without weights; Meta’s only open-weight Muse model is the 30B Muse Glimmer (10 Aug, Apache 2.0).
- 2026-09-06 → 10-12 The asking: users press Meta on Muse Spark weights (“'Soon' and a month later theres no update”; “So, Meta is definitely not open sourcing their models anymore, are they?” — @jmbollenbacher).
- 2026-10-02 Framework: Meta revises its Superintelligence Scaling Framework, adding open-weight considerations, after White House commitments.
- 2026-10-15 Release: three days after Muse Wonder, the weights go out under the Muse Model License, with Mac Studio (MLX) support and custom Helion GPU kernels for SGLang.
Impressions
- tk — too new: no corpus record or independent reports yet. Watch for: whether the extra safety post-training reads as refusal-heavy in use (its StrongREJECT figure is far below the other open models listed); fine-tuned derivatives; how it compares in character with the served Muse Spark versions.