Open-weight model

Muse Spark Open

Muse Spark Open is the most capable open-weight model from a US lab, and is part of our commitment to giving developers control over how and where they choose to build.

Documentation

Where to deploy Muse Spark Open

AMD
baseten
Fireworks
Cloud
groq
Hugging Face
vLLM
Microsoft
Modal
NEBiUS
NVIDIA
openrouter
SGLang
together.ai
Unsloth

Meet Muse Spark Open

agent / muse-spark-open
Reviewing the change and validating the build

Muse Spark Open writes, reviews and ships code with fewer steps and lower latency. Whether you are building coding agents or using AI as a development partner, it is competitive with the best.

Muse Spark

Introducing Muse Spark Open

Muse Spark Open is derived from the Muse Spark model we released in August, with additional safety post-training to make it more robust against misuse as a standalone model without system guardrails.

Read more

Muse Spark Open benchmarks

CategoryBenchmarkMuse Spark OpenInklingNemotron 3 UltraBeamMistral Large 4Qwen3.8-2.4TGLM 5.3Kimi K3
AgenticJobBench61.831.127.2––52.761.454.3
MCP-Atlas89.276.063.178.7–84.584.282.3
GDPval-AA v2140110791016–1424161316531537
HLE-Diamondno tools26.111.28.9––17.316.422.2
Knowledge & ReasoningSciCode58.547.040.349.754.254.159.059.5
AA-Omniscience252013-541420
Terminal-Bench 2.1Terminus 2 agent76.455.153.980.1–82.083.985.0
DeepSWE 1.1Best available harness60––4462576768
CodingZeroBench50.014.0Not supportedNot supported–Not supportedNot supported41.0
BabyVision80.234.2Not supportedNot supported–Not supportedNot supported85.7
Multimodal (with tools)SimpleVQA73.652.3Not supportedNot supported–Not supportedNot supported71.9