Model
Model explorer

Muse Spark Open

OPEN
Meta · Muse Spark family · released Oct 15, 2026

Meta's open-weight release of Muse Spark, October 15, 2026: an 816B-total / 42B-active natively multimodal MoE (text, image, video and audio in, text out) with a 1M-token context, shipped in BF16 and NVFP4 under the Muse Model License and served by third-party hosts (Baseten, Fireworks, Groq, Modal, Nebius, OpenRouter, Together); Meta sets no first-party price. Meta's table runs it at xhigh effort against open-weight peers: it leads JobBench (61.8) and MCP-Atlas (89.2) but trails GLM 5.3 and Kimi K3 on DeepSWE 1.1 (60) and Terminal-Bench 2.1 (76.4). NOT FILED: Artificial Analysis Intelligence Index (39) has no corpus column; GDPVal-AA v2 (1401) is not filed because the corpus gdpval-aa column and AA's re-based v2.x launch figures are on different scales; HLE-Diamond no-tools (26.1) is a subset, not the full HLE; AA Omniscience (25), ZeroBench, BabyVision, SimpleVQA, WorldVQA, IntPhys2 and VideoMME v2 have no matching column; FORTRESS and StrongREJECT are safety metrics.

ReasoningCodingVisionFunction callingTool useAgentic
2595.7
Elo · rank #32
Parameters
816B
Active params
42B (MoE)
Context
1M tokens
Architecture
Sparse MoE transformer, 816B total / 42B active (256 experts, top-8 routing, 62 layers, GQA with gated attention) plus a 1.8B perception encoder; 1M-token context; reasoning effort low to xhigh
License
Muse Model License (free use with attribution; Model-as-a-Service providers above $20M annual revenue need a separate agreement with Meta)
Languages
—
API price (in/out)
No hosted API
Modalities
text · vision · audio · video
Benchmark results
Bar shows position within the tracked field; marker = field best
DeepSWECoding60.0%#25
best: Muse Wonder Full · 79.0%
GPQA DiamondReasoning89.0%#43
best: GPT-6 Astra · 96.0%
best: Muse Spark 1.2 · 25.4%
Humanity's Last ExamReasoning61.0%#5
best: Kimi K3.1 Pro · 70.0%
JobBenchAgents61.8%#7
best: Muse Wonder Full · 75.3%
MCP AtlasAgents89.2%#1
best: this model · 89.2%
OSWorld-VerifiedAgents80.4%#9
best: Claude Fable 5 · 85.0%
SciCodeCoding58.5%#2
best: Step 5 Preview · 58.9%
best: Step 5 Preview · 63.6%
Terminal-Bench 2.0Coding76.4%#25
best: DeepSeek-V4.1-Flash · 90.6%
Terminal-Bench 4.0Agents6.6%#40
best: Claude Sonnet 5.5 · 70.6%
Toolathlon (Verified)Agents72.8%#11
best: GLM-5.3-Flash · 78.4%
Vals Finance AgentAgents61.3%#3
best: Gemini 4 Argon · 65.4%
Run it locally
VRAM @ Q4
700 GB
VRAM @ FP16
2000 GB
Fits on (Q4)
Multi-node cluster required
Throughput data unavailable.
Quantizations
NVFP4
Fine-tune it
Conditional / custom
QLoRA554.9 GB4× H200 141GB
LoRA1738.1 GBbeyond 8× B200
Full fine-tune13096.8 GBbeyond 8× B200
QLoRA SFT on ~10k samples ≈ $21.72 (4× H200 141GB)
API price weights · each benchmark row carries its own source badge (see methodology)