Meta and Nvidia Launch Competing Open AI Models Simultaneously

On October 15, 2026, Meta released the open model Muse Spark Open, calling it the most powerful American development with open weights. The system is tailored for coding and agent work, understanding text, images, video, and audio. It features 816 billion parameters, with 42 billion active simultaneously, and a context window of 1 million tokens. On the same day, Nvidia introduced a competing open model, Nemotron 3.5 Super VL. Both releases occurred simultaneously, targeting the same market segment. Meta confirmed the specifications on X. The model is designed for long-term tasks, complex software engineering, and multi-step reasoning. APIs are available on Baseten, Groq, and Together AI, with weights on Hugging Face and in the clouds of Amazon, Google, and Microsoft. Downloaded weights can be run on personal equipment. Meta optimized the model for two Apple M5 Ultra Mac Studios through MLX. It is compatible with SGLang and vLLM, operating on NVIDIA, AMD, and Intel hardware.

According to the Artificial Analysis Intelligence Index 4.3.2, Muse Spark Open scores 39 points in xhigh mode. Among American models with available weights, this is the best result, while Nemotron 3.5 Super VL scores 26 points. Chinese developments are ahead: GLM 5.3 (max) scores 45, Kimi K3 (max) scores 44. Nvidia built Nemotron 3.5 Super VL based on the mixture of experts principle. Out of 120 billion parameters, 12 billion work simultaneously, making the model faster and cheaper. It is designed for agent tasks: the AI performs a sequence of actions, follows instructions, calls external programs, and analyzes images. Nvidia claims that the model matches the accuracy of its largest model with step-by-step reasoning but is about five times smaller in terms of parameters. The company calls it the fastest open agent model in its class: on the same tasks, it processes up to five times more data per unit of time.

The model can be downloaded on Hugging Face, including in the compressed NVFP4 version, where numbers are recorded in four bits. The context window is 1 million tokens, and it can be run on a single graphics card, including DGX Spark and H200. Alongside it, the Bolt family was released, compact models for running on user devices without the cloud. Analysts note that Meta has entered a scene where roles have changed: in August, publications wrote that leading American developers increasingly choose closed systems, while Chinese companies have become the main suppliers of open weights. Meta itself opened Muse Spark for closed API preview only in April. Therefore, the October release seems like a counter move. The term open requires clarification: the weights are released under the Muse Model License with limited commercial terms.