ModelsArchitectures & capability
Meta releases open weights for Muse Spark Open, an 816B-parameter mixture-of-experts model
Six months after shipping Muse Spark as a closed model, Meta has published weights for an 816B-parameter mixture-of-experts version that activates 42B parameters per token, handles text, image, video and audio, offers 1M-token context, and targets coding and agent work.
From the huggingface.co article of October 15, 2026 and cited sources: This reverses Meta's April decision to release Muse Spark as a proprietary model, which broke from its open Llama approach. [2] [4] [5] [9]
SOURCES[2] The Next Web: Meta Muse Spark, MSL's first model thenextweb.com[4] Bloomberg Law: Meta Debuts First AI Model From New… news.bloomberglaw.com[5] PyTorch: Muse Spark 1.2 Is Now Open Weights pytorch.org
AT Pulse research deskPublished · Updated · 5 min read · AI-assisted analysisAI-assisted, not human-edited. Methodology Corrections Who owns this site
- What happened
- This reverses Meta's April decision to release Muse Spark as a proprietary model, which broke from its open Llama approach. [2] [4] [5] [9]
- Original source
- huggingface.co article, company source
- Brief by
- AT Pulse research desk AI-assisted; publisher and contact on the desk page
- When
- October 15, 2026 article published
- Covered
- Oct 16 edition
- Evidence
- 9 sources across 9 sites
- 1Hugging Face: Muse-Spark-Open-816B-A42Bhuggingface.co
- 2The Next Web: Meta Muse Spark, MSL's first modelthenextweb.com
- 3Artificial Analysis on Xx.com
- 4Bloomberg Law: Meta Debuts First AI Model From New…news.bloomberglaw.com
- 5PyTorch: Muse Spark 1.2 Is Now Open Weightspytorch.org
- 6OrcaRouter: Muse Spark Open: open weights out, benchmarks…orcarouter.ai
RisingHow it’s measured
Hacker News · 8 likes · 2 comments
Reach blends public engagement (views, likes, comments, shares, stars) from the platforms where the story or its original source lives: YouTube, Hacker News, GitHub, Substack and similar. A video and the article written from its transcript share one score, taken from the video. Platforms with no data are left out rather than counted as zero, and the age of a story never lowers its score.
This reverses Meta's April decision to release Muse Spark as a proprietary model, which broke from its open Llama approach. Until now, Meta's only open Muse model was the small dense Muse Glimmer, built for on-device use. Muse Spark Open is the first full-size model in the family with public weights. [2] [4] [5] [9]
So far, Artificial Analysis has the only third-party measurement. It scores the model 39 on its Intelligence Index. That makes it the top US open-weights model, one point behind the closed Muse Spark 1.2 and 22 points ahead of Muse Glimmer. The model card's higher figures, such as Terminal-Bench 2.1 at 76.4 and OSWorld Verified at 80.4, are vendor-reported. Each was run in its own agent harness, which makes comparison with other models harder. [3] [1] [6]
Teams can try it before self-hosting. It launched on seven inference providers and on major cloud model catalogs, in BF16 and NVFP4 weights plus partner quantizations. The license allows commercial use and modification, but API providers earning more than $20M over 12 months must notify Meta and sign a separate agreement. Meta's own documents give inconsistent hardware figures, and full 1M context needs datacenter-class memory. [1] [7] [8] [6]
Meta says the open version got extra safety post-training so it holds up without system guardrails. It also reports a 99.2% jailbreak refusal rate on an internal cybersecurity benchmark. Even so, Meta still recommends system-level guardrails, and its detailed safety evaluation report has not been published. [5] [1] [6]
What's established: a frontier-scale model with open weights, usable commercially, is available from several hosts now, and one independent index ranks it near Meta's closed model.
Read the full assessment
The implication is cheaper experimentation and more control for teams, but they should test it on their own workloads first.
Executive brief
Meta, which made Muse Spark a closed model in April, has now published weights for an 816B-parameter version. Muse Spark Open is a mixture-of-experts (MoE) model that activates 42B parameters per token. It has a 1M-token context window, accepts text, image, video and audio, and Meta built it for coding and agent work. The only third-party measurement so far is from Artificial Analysis, which scores it 39 on its Intelligence Index. That ranks it the top US open-weights model and one point below the closed Muse Spark 1.2 (40). Meta has not yet published its detailed safety report.
What changed and event timeline
Muse Spark 1.2 released
Meta's August update is the release the open version builds on, according to the.
Muse Spark launches closed
Meta Superintelligence Labs' first model went out as proprietary, a break from the open Llama strategy.
More detail
Gated repository appears quietly
The Hugging Face repo went up gated and unannounced and stayed quiet for a month ().
Weights publicly released
Weights went out under the Muse Model License in BF16 and NVFP4 formats, with hosted access through seven inference providers (;).
First third-party score
Artificial Analysis gave it 39, which is 22 points above Meta's 30B dense open model, Muse Glimmer (17) ().
Capabilities and access
- Model: Muse-Spark-Open-816B-A42B. A Transformer MoE with a perception encoder, 62 layers, 256 experts (8 active per token) and grouped-query attention.
- Formats: Meta's own BF16 and NVFP4 weights. Partners add FP8 and MXFP4 (AMD), INT4 (Intel) and roughly 2-bit GGUF builds (Meta).
- Hosted access: Baseten, Fireworks, Groq, Modal, Nebius, OpenRouter and Together. It is also on Google Cloud Model Garden, SageMaker and Microsoft Foundry (Meta research blog).
Read the full section
- Model: Muse-Spark-Open-816B-A42B. A Transformer MoE with a perception encoder, 62 layers, 256 experts (8 active per token) and grouped-query attention. It reads text, image, video and audio, writes text, and covers 100+ languages.
- Formats: Meta's own BF16 and NVFP4 weights. Partners add FP8 and MXFP4 (AMD), INT4 (Intel) and roughly 2-bit GGUF builds (Meta).
- Hosted access: Baseten, Fireworks, Groq, Modal, Nebius, OpenRouter and Together. It is also on Google Cloud Model Garden, SageMaker and Microsoft Foundry (Meta research blog).
Technical analysis for researchers and developers
- Lineage: PyTorch says the model shares its base with Muse Spark 1.3.
- Kernels: It ships with Helion kernels (RoPE, RMSNorm, quantization).
- Fine-tuning: TorchTitan supports SFT, LoRA and RL out of the box (PyTorch).
Read the full section
- Lineage: PyTorch says the model shares its base with Muse Spark 1.3. The same post's title calls it "Muse Spark 1.2", and it says the open version received extra safety post-training so it holds up without system guardrails.
- Kernels: It ships with Helion kernels (RoPE, RMSNorm, quantization). PyTorch reports 1.03–3.95× kernel-level speedups on GB300 and a 3% end-to-end gain on AMD MI350 (PyTorch).
- Fine-tuning: TorchTitan supports SFT, LoRA and RL out of the box (PyTorch).
- Speculative decoding: Meta's draft models (DFlash V2, DSpark) are claimed to raise throughput 2.9–3.7× (Meta).
- Memory: OrcaRouter estimates the KV cache alone needs about 266GB in BF16 at full 1M context (OrcaRouter).
Claims and evidence
- "Most capable US open-weights model": this is Meta's claim. Artificial Analysis's ranking is consistent with it.
- Benchmark scores: the model card reports Terminal-Bench 2.1 76.4, GPQA Diamond 89 and OSWorld Verified 80.4. These are vendor-reported, and each was run in its own agent harness, which limits comparison with other models.
- Safety claims: the card says the model refused 99.2% of jailbreaks on an internal cybersecurity benchmark. No independent replication exists.
Read the full section
- "Most capable US open-weights model": this is Meta's claim. Artificial Analysis's ranking is consistent with it. That ranking is limited to US labs, so it says nothing about non-US open models.
- Benchmark scores: the model card reports Terminal-Bench 2.1 76.4, GPQA Diamond 89 and OSWorld Verified 80.4. These are vendor-reported, and each was run in its own agent harness, which limits comparison with other models.
- Safety claims: the card says the model refused 99.2% of jailbreaks on an internal cybersecurity benchmark. No independent replication exists.
Context and prior work
Meta dropped open releases in April when it shipped Muse Spark as a closed model. Before this release, Meta's open Muse model was the small dense Muse Glimmer, aimed at on-device use through ExecuTorch. Muse Spark Open is the first full-size model in the family with open weights. Meta evaluated it under its "Meta Superintelligence Scaling Framework".
Read the full section
Meta dropped open releases in April when it shipped Muse Spark as a closed model. Some commentators expected open weights to return. Before this release, Meta's open Muse model was the small dense Muse Glimmer, aimed at on-device use through ExecuTorch. Muse Spark Open is the first full-size model in the family with open weights. Meta evaluated it under its "Meta Superintelligence Scaling Framework".
Limitations, safety and contested findings
- Safety report not out: Meta says the model met its framework's release thresholds.
- Uneven strengths: in Meta's own numbers, OrcaRouter finds agentic tasks are strong and reasoning benchmarks weaker.
- Hardware figures don't line up: the model card lists 700GB of VRAM for NVFP4 and 2TB for BF16. Meta's developer blog gives weight sizes of about 530GB and 1.6TB.
Read the full section
- Safety report not out: Meta says the model met its framework's release thresholds. The detailed evaluation report is still listed as "available soon".
- Uneven strengths: in Meta's own numbers, OrcaRouter finds agentic tasks are strong and reasoning benchmarks weaker.
- Hardware figures don't line up: the model card lists 700GB of VRAM for NVFP4 and 2TB for BF16. Meta's developer blog gives weight sizes of about 530GB and 1.6TB. The gap may be weights versus runtime memory, but neither source says so.
- Version labels disagree: the 1.2 vs 1.3 question above is unresolved.
- Guardrails still advised: Meta recommends system-level guardrails even with the extra safety training.
Business and practitioner implications
- License: the model can be used commercially and modified. Large API resellers should read the full text.
- Deployment: Meta demonstrated it running locally on two M5 Ultra Mac Studios. Full 1M context, though, needs datacenter-class memory.
- Hosting choice: with seven inference providers available from day one, teams can pilot it hosted before deciding whether to run it themselves.
Read the full section
- License: the model can be used commercially and modified. Attribution is required, and any company that serves it to others through an API and earns more than $20M over 12 months must notify Meta and sign a separate agreement (OrcaRouter summary of Section 3). Large API resellers should read the full text.
- Deployment: Meta demonstrated it running locally on two M5 Ultra Mac Studios. Full 1M context, though, needs datacenter-class memory.
- Hosting choice: with seven inference providers available from day one, teams can pilot it hosted before deciding whether to run it themselves.
- Before adopting: run your own benchmarks on your own workloads, and wait for Meta's safety report.