Muse Spark
Muse Spark is a multimodal artificial intelligence reasoning model that combines visual understanding with text-based reasoning. It supports tool invocation, visual chain-of-thought reasoning, and multi-agent coordination. The model is designed to handle complex cross-modal tasks and can serve as a reasoning core within multi-agent systems, making it applicable to scenarios that require integrated perception and logical inference.
Core Facts
Muse Spark Open can be called on the day of release through platforms such as Baseten, Fireworks AI, Groq, Modal, Nebius, OpenRouter and Together AI.
Muse Spark Open uses an MoE architecture, with 816B total parameters and 42B activated parameters.
Muse Spark can be used as a backend via OpenCode Go.
Timeline (last 90 days)
According to the video, Big Pickle, Mimo, Muse Spark, and the Nemotron series may use users' prompts to improve their models, so passwords, API keys, or confidential client code should not be pasted in.
The key decisions for running Muse Spark include the quantization scheme (FP8/INT8/4-bit), the inference framework (vLLM, TensorRT-LLM, SGLang), and the parallelism strategy (Tensor Parallel, Pipeline Parallel).
Muse Spark's license is not an unrestricted open-source license such as Apache 2.0, but may include custom community license conditions such as a monthly active user threshold, commercial use restrictions, or prohibitions in specific scenarios.
Meta officially acknowledges that Muse Spark still lags on complex tasks requiring deep chain-of-thought reasoning and in some multilingual scenarios.
Muse Spark approaches but does not surpass the current strongest closed-source models on most benchmarks, and performs outstandingly on code generation and mathematical reasoning.
A single machine with 8×H100 (640GB of VRAM) can barely load Muse Spark under FP8, but actual inference with a 1M context requires more aggressive sharding or quantization strategies.
Meta officially released the open weights of Muse Spark on October 15, 2026.
Meta had previously said it would open the weights of Muse Spark, but when Meta released the closed-source Muse Wonder frontier model on October 13, the open weights originally slated for delivery on September 2 had still not materialized.
Muse Spark Open uses an MoE architecture, with 816B total parameters and 42B activated parameters.
Muse Spark Open can be called on the day of release through platforms such as Baseten, Fireworks AI, Groq, Modal, Nebius, OpenRouter and Together AI.
22 more timeline events
All Facts (20)
Muse Spark Open can be called on the day of release through platforms such as Baseten, Fireworks AI, Groq, Modal, Nebius, OpenRouter and Together AI.
90%VerifiedMuse Spark Open uses an MoE architecture, with 816B total parameters and 42B activated parameters.
85%VerifiedMuse Spark can be used as a backend via OpenCode Go.
85%VerifiedMuse Spark is a model specifically designed for coding assistants and intelligent agents.
75%VerifiedMeta publicly stated on August 10 that it would release the weights of the Muse Spark model as open source.
65%UnverifiedMuse Spark是NVIDIA Omniverse生态系统的一部分
80%UnverifiedMuse Spark聚焦于降低3D内容创作的门槛，利用生成式AI技术让普通用户能完成此前需要专业技能的创作任务
80%Partially VerifiedThe 816B-parameter Muse Spark Open occupies about 493GB of memory after NVFP4 quantization, while FP16 storage would require more than 1.6TB.
80%UnverifiedEach model has an independent quota pool, so high traffic on Muse Wonder will not crowd out Muse Spark's channel, and vice versa.
60%UnverifiedAccording to the video, Big Pickle, Mimo, Muse Spark, and the Nemotron series may use users' prompts to improve their models, so passwords, API keys, or confidential client code should not be pasted in.
50%UnverifiedMuse Spark's license is not an unrestricted open-source license such as Apache 2.0, but may include custom community license conditions such as a monthly active user threshold, commercial use restrictions, or prohibitions in specific scenarios.
50%UnverifiedMeta officially acknowledges that Muse Spark still lags on complex tasks requiring deep chain-of-thought reasoning and in some multilingual scenarios.
50%UnverifiedA single machine with 8×H100 (640GB of VRAM) can barely load Muse Spark under FP8, but actual inference with a 1M context requires more aggressive sharding or quantization strategies.
50%UnverifiedMuse Spark approaches but does not surpass the current strongest closed-source models on most benchmarks, and performs outstandingly on code generation and mathematical reasoning.
50%UnverifiedThe key decisions for running Muse Spark include the quantization scheme (FP8/INT8/4-bit), the inference framework (vLLM, TensorRT-LLM, SGLang), and the parallelism strategy (Tensor Parallel, Pipeline Parallel).
50%UnverifiedMeta did not publish official requests-per-minute (RPM) or tokens-per-minute (TPM) limits for Wonder, continuing the black-box approach from the earlier Muse Spark launch.
50%UnverifiedAfter Muse Spark was released, it quietly became a primary tool, and someone on Hacker News said it migrated 90% of their programming work away from Codex.
50%UnverifiedMuse is a personal AI assistant launched by Meta, running on the Muse Spark model, currently limited to the United States and available on iOS, Android, web, and WhatsApp
50%UnverifiedMeta's Muse is a consumer-grade product oriented toward personal scenarios, using the Muse Spark model.
50%UnverifiedMuse is powered by Meta's in-house muse spark 1.3 model.
50%