Skip to content
AIMODELRACE
Meta Superintelligence LabsLanguage modelOpen weightsNew

Muse Spark Open

The most capable US open-weights model: multimodal in, 1M context, runs on your own hardware.

Compare head-to-headRace it in 3D

Overall Intelligence

#20

of 56
8.2 pts behind #1

0

races led

0

weeks at #1

80.5

score

Released
Oct 15, 2026
4 days ago
Context
1M
64K max output
Price
$0.6 / $2.4
per 1M in/out · $1.05 blended
Speed
125 tok/s
1.4s to first token
Modalities
text · image · audio · video
outputs text
Weights
Open
~816B MoE (~42B active)

The 30-second version

The most capable US open-weights model. Muse Spark Open takes text and images, handles 1M tokens of context and runs on your own hardware.

Key takeaways
  • 1Leads the open-weights pack from US labs in our current snapshot.
  • 2Self-host for privacy, customisation and fine-tuning.
  • 3Multimodal input and a 1M-token context.
  • 4Behind the closed frontier by several points on the hardest tasks.

What you need to know

Purpose

Give developers a frontier-adjacent model they can own, fine-tune and deploy anywhere.

What Meta optimised for

Open, efficient and fine-tunable, with permissive-enough licensing for most companies.

Unique strengths

  • Open weights with multimodal input.
  • Efficient inference.
  • Fine-tuning friendly.
Insight
How it differs from rivals. Versus DeepSeek V4 Pro: easier to deploy, multimodal, slightly weaker at competitive programming. Versus Kimi K3: much smaller and cheaper to run.

Where it ranks

Current snapshot · illustrative scores · click any race to watch it

CodingHard ReasoningMathScienceWriting & LanguageMultimodalTool Use & AgentsLong ContextTrust & FactualitySpeed & Cost

Spark OpenBest on each axis

Overall Intelligence over time

#1 frontierSpark OpenQwen3.6-MaxHaiku 5.5
55.060.065.070.075.080.085.090.02025Apr ’25Jul ’25Oct ’252026Apr ’26Jul ’26Oct ’26

Best use cases

01

Private deployments

Regulated data that cannot leave your infrastructure.

02

Fine-tuned specialists

Domain models for legal, medical or support use.

03

Edge and on-prem

Running close to users or data.

Strengths

  • Openness

    Weights available for self-hosting and customisation.

Weaknesses

  • Frontier gap

    Several points behind the closed leaders on hard reasoning.

When to choose it, and when not to

Spark OpenoverV4 Pro

You need image input or a simpler deployment footprint.

Pick something else when…

  • You need the absolute best quality and do not need to self-host.

Update history

The distilled version first; open any release for the deeper breakdown.