Close Menu
  • AI
  • Content Creation
  • Tech
  • Robotics
AI-trends.todayAI-trends.today
  • AI
  • Content Creation
  • Tech
  • Robotics
Trending
  • A Group of Researchers from Columbia and CertiK Introduce TensorLean for Formal Verification of GPU Kernels
  • Collectively AI Discloses Immediate Leak After Coding Agent Uncovered Credential – Unite.AI
  • Dragos Finds Immediate Leaks and Stolen Keys in AI Provide Chains
  • Lean and Mathlib Communities Certify Claude-Assisted Riemann Proof – Unite.AI
  • AWS Routes HyperPod Inference Throughout Areas From One Endpoint – Unite.AI
  • DailyAI Briefing: A Lean-checked Riemann proof, plus 4 extra
  • Anthropic Shares Drop 8.5% in Second Session Throughout AI Selloff – Unite.AI
  • OpenAI Ships GPT-Stay-1.1 for Voice Brokers, Cuts Idle Value to $0.02/Min
AI-trends.todayAI-trends.today
Home»Tech»Meta Releases Muse Spark Open: 816B Open-Weight MoE With 1M Context

Meta Releases Muse Spark Open: 816B Open-Weight MoE With 1M Context

Tech By Gavin Wallace16/10/20265 Mins Read
Facebook Twitter LinkedIn Email
Mistral Launches Agents API: A New Platform for Developer-Friendly AI
Mistral Launches Agents API: A New Platform for Developer-Friendly AI
Share
Facebook Twitter LinkedIn Email

Meta Superintelligence Labs has launched Muse Spark Open, the open-weight model of its Muse Spark mannequin. Muse Spark Open is an 816B-parameter Combination-of-Specialists mannequin that prompts 42B parameters per token. It takes textual content, picture, video and audio enter, returns textual content, and holds a 1M-token context window. Meta workforce claims it to be essentially the most succesful US open-weight mannequin thus far. For builders, the actual information is deployment: the NVFP4 construct runs on 2 Apple M5 Extremely Mac Studios.

TL;DR

  • Measurement: 816B complete / 42B energetic parameters (256 specialists, 8 energetic per token); 1M-token context.
  • Runs on: 2 Mac Studios (M5 Extremely, 512 GB every) through MLX; 4 Blackwell Extremely GPUs or 8 H200s with NVFP4; about 300 GB setups through Unsloth 2-bit GGUF.
  • Efficiency: Robust on agentic device use; mid-pack amongst giant open fashions on coding.
  • Greatest: MCP-Atlas 89.2, the highest rating in Meta’s comparability desk.
  • Worst: Terminal-Bench 2.1 76.4, behind Kimi K3 (85.0), GLM 5.3 (83.9) and Qwen3.8 (82.0).

What’s Muse Spark Open?

Muse Spark Open is a natively multimodal MoE mannequin from Meta, constructed for coding and long-horizon agentic work. Weights ship in BF16 and NVFP4 on Hugging Face. The open model acquired additional security post-training. Meta workforce states that this makes it extra sturdy with out system-level guardrails.

How does the structure work?

The model card lists 62 layers and a hidden dimension of 8192. Every MoE layer holds 256 specialists, and top-k routing prompts 8 per token. Consideration is grouped-query, with 128 question heads and 16 KV heads. A separate 1.8B-parameter Notion Encoder handles visible enter. The vocabulary has 202,048 tokens, and coaching knowledge covers greater than 100 languages.

Reasoning depth is about with reasoning_effort: low, medium, excessive or xhigh. The default is xhigh.

How does Muse Spark Open carry out on benchmarks?

Meta compares it with Kimi K3, GLM 5.3, Qwen3.8-2.4T-A95B and others at most reasoning effort. Its strongest outcomes are on device use and information reliability:

  • MCP-Atlas: 89.2, forward of Qwen3.8 (84.5) and GLM 5.3 (84.2).
  • JobBench: 61.8, narrowly forward of GLM 5.3 (61.4).
  • HLE-Diamond, no instruments: 26.1, versus 22.2 for Kimi K3.
  • AA Omniscience: 25, the highest rating in Meta’s desk.
  • ZeroBench with instruments: 50.0, versus 41.0 for Kimi K3.

Coding is the weak spot. On Terminal-Bench 2.1 with the Terminus 2 harness, it scores 76.4. Kimi K3 reaches 85.0, GLM 5.3 scores 83.9 and Qwen3.8 scores 82.0. Its Synthetic Evaluation Intelligence Index is 39. That sits behind GLM 5.3 (45), Kimi K3 (44) and Qwen3.8 (40). So “most capable US” holds, however Chinese language open fashions nonetheless lead general.

Some rival scores are self-reported; the evaluation methodology lists every supply.

(operate(){var f=doc.getElementById(‘mtp-mso-frame’);window.addEventListener(‘message’,operate(e){if(e.knowledge&&e.knowledge.mtpEmbed===’muse-spark-open’&&e.knowledge.top){f.fashion.top=e.knowledge.top+’px’;}});})();

What {hardware} does Muse Spark Open want?

Per Meta’s developer guide, BF16 weights are about 1.6 TB. NVFP4 quantizes solely the specialists and lands close to 530 GB. The mannequin card recommends no less than 700 GB of VRAM for NVFP4 and a couple of TB for BF16.

  • Information heart: NVFP4 on 4 Blackwell Extremely GPUs or 8 H200s.
  • Desktop: 2 M5 Extremely Mac Studios with 512 GB every, linked over Thunderbolt 5 with RDMA. MLX tensor parallelism splits each layer throughout each machines.
  • Workstation: Unsloth Dynamic GGUFs right down to 2-bit slot in about 300 GB of mixed reminiscence.

The place can builders run it?

Day-one engines embrace SGLang, vLLM, TensorRT-LLM, Hugging Face Transformers, MLX, llama.cpp and Unsloth. Hosted APIs are dwell on Baseten, Fireworks AI, Groq, Modal, Nebius, OpenRouter and Collectively AI. OpenRouter lists $1.25 per 1M enter tokens and $4.25 per 1M output tokens.

Two draft fashions pace up decoding: DFlash V2 for SGLang and DSpark for vLLM. Meta studies 3.7x single-request throughput with DFlash V2 on AMD MI355X. High-quality-tuning works with TorchTitan, Hugging Face TRL and NVIDIA NeMo AutoModel.

What does the Muse Mannequin License permit?

The Muse Model License permits use, fine-tuning, distillation and redistribution. 2 situations matter most:

  • Merchandise should show “Built with Muse Spark Open.”
  • Mannequin-as-a-Service suppliers above $20M in 12-month income want a separate settlement with Meta.

Inner use, analysis and powering your individual product options are exempt from the income clause.

Muse Spark Open vs Kimi K3 vs GLM 5.3 vs Qwen3.8

Function Muse Spark Open Kimi K3 GLM 5.3 Qwen3.8-2.4T-A95B
Whole params 816B 2.8T 753B 2.4T
Lively params 42B 104B ~40B 95B
Context 1M 1,048,576 1,048,576 262,144 native, as much as 1,010,000
Enter Textual content, picture, video, audio Textual content, picture, video Textual content Textual content
License Muse Mannequin License Kimi K3 License glm-5.3 license qwen3.8-max license
Launched weights BF16, NVFP4 MXFP4 (QAT) BF16, FP8 BF16
AA Intelligence Index (via Meta) 39 44 45 40
Terminal-Bench 2.1, Terminus 2 (via Meta) 76.4 85.0 83.9 82.0
MCP-Atlas (via Meta) 89.2 82.3 84.2 84.5
Toolathlon Verified (vendor) 72.8 76.5 73.0 72.5
HLE with instruments (vendor) 61.0 56.0 62.5 56.2
DeepSWE 1.1 (vendor, personal harness) 60 67.5 66.9 56.6
Desktop setup 2x M5 Extremely Mac Studio Not disclosed Not disclosed Not disclosed

Key Takeaways

  • 816B complete, 42B energetic, 1M context, textual content/picture/video/audio enter.
  • Leads Meta’s open-model desk on MCP-Atlas (89.2) and JobBench (61.8).
  • Trails Kimi K3, GLM 5.3 and Qwen3.8 on Terminal-Bench 2.1 (76.4).
  • NVFP4 construct runs on 2 M5 Extremely Mac Studios through MLX.
  • Free for many; MaaS suppliers above $20M income want a deal.


Take a look at the model weights, research blog, developer guide and evaluation methodology. All credit score goes to the researcher of this challenge. Additionally, be happy to comply with us on Twitter and don’t overlook to hitch our 150k+ML SubReddit and Subscribe to our Newsletter. Wait! are you on telegram? now you can join us on telegram as well.

The put up Meta Releases Muse Spark Open: 816B Open-Weight MoE With 1M Context appeared first on MarkTechPost.

ar met meta muse x
Share. Facebook Twitter LinkedIn Email
Avatar
Gavin Wallace

Related Posts

A Group of Researchers from Columbia and CertiK Introduce TensorLean for Formal Verification of GPU Kernels

22/11/2026

OpenAI Ships GPT-Stay-1.1 for Voice Brokers, Cuts Idle Value to $0.02/Min

21/11/2026

Meet Pondering Machines Inkling 2 Mini: 26B Looped MoE Runs on a 32GB Mac

21/11/2026

Baseten Releases Bison-1 and Groundhog-1, Topping Choice Mannequin Index

21/11/2026
Top News

Bryan Johnson is going to die

Teachers try to use AI in a way that works for them

AI and Water: Your Thinking is Wrong

The director of a raunchy, 3-hour Dracula film says that AI is slimy and gross. The Director of a Raunchy 3-Hour Dracula Movie Says AI Is Gross and Slimy

Grok Still Hosts Sexualized Deepfakes Of Famous Women

Load More
AI-Trends.Today

Your daily source of AI news and trends. Stay up to date with everything AI and automation!

X (Twitter) Instagram
Top Insights

Feyn AI releases SQRL, which is a Text to SQL model family that inspects the database before writing a query

19/07/2026

The Top 10 Personal Branding Courses – Detailed Analysis

01/08/2025
Latest News

A Group of Researchers from Columbia and CertiK Introduce TensorLean for Formal Verification of GPU Kernels

22/11/2026

Collectively AI Discloses Immediate Leak After Coding Agent Uncovered Credential – Unite.AI

22/11/2026
X (Twitter) Instagram
  • Privacy Policy
  • Contact Us
  • Terms and Conditions
© 2026 AI-Trends.Today

Type above and press Enter to search. Press Esc to cancel.