Meta, NVIDIA and Reflection release open weights for very different machines

Three releases in two days

Reflection released Beam’s weights on October 16, following open-model launches from Meta and NVIDIA the previous day. The releases offer very different routes to running an agent: an enormous multimodal model on a pair of high-memory desktops, a model compressed for a much smaller weight footprint, and a new coding model released with an open tooling stack.[6][1][3]

All three put weights in developers’ hands. Their practical requirements depend on the exact checkpoint, its license and the software serving it. A parameter count alone leaves much of that story out.

Meta’s desktop demonstration has a terabyte of memory

Muse Spark Open has 816 billion parameters, with 42 billion active for each token, and a one-million-token context window. Meta built the multimodal model for coding and agentic work and says it can run through engines including MLX, SGLang and vLLM.[1]

Meta’s Apple demonstration uses two M5 Ultra Mac Studios, each offering up to 512GB of unified memory. The company describes the pair as a desktop system with a terabyte of memory and reports on-device workflows that include changing an iOS app and working with video in DaVinci Resolve. These are Meta’s demonstrations of a particular setup.[1]

The public model card gives format-specific requirements: at least 700GB of VRAM for NVFP4 and 2TB for BF16. Those author requirements describe different formats from one another; the desktop demonstration uses a particular optimized setup.[2]

Activating a fraction of the experts reduces work for an individual token. The remaining weights still require storage, while context caches and serving software also consume memory. Meta’s example shows why the number of active parameters gives an incomplete picture of the hardware bill.

NVIDIA starts with a smaller weight footprint

Nemotron 3.5 Bolt 74B-A9B uses 74 billion total parameters and nine billion active parameters. NVIDIA’s 24GB release applies native two-bit quantization and accepts text, images and video. Its model card provides a specific vLLM serving recipe.[4]

The 24GB figure describes the packed model. Context caches, runtime overhead and other applications still affect the machine’s usable budget. The published recipe and context setting matter when turning that weight footprint into a working deployment.[4]

NVIDIA also released Nemotron 3.5 Super VL: a larger vision-language model with 120 billion total parameters, 12 billion active parameters and support for context up to one million tokens. Its BF16 checkpoint is intended for full-precision work and customization, with separate optimized deployment options.[5]

Reflection releases the model and the surrounding tools

Beam is a 501-billion-parameter mixture-of-experts model with 23 billion active parameters, built for coding, reasoning and agentic workloads. Reflection published BF16, FP8 and NVFP4 formats and describes launch support across serving engines, hosted platforms and agent harnesses.[7][6]

The accompanying Mirror CLI lets developers inspect an agent’s activity, approve tool calls and resume work. Reflection is also releasing integrations and customization workflows around the model, extending the launch beyond a checkpoint download.[6]

The licenses differ as well. Reflection’s weights use Apache 2.0. NVIDIA’s two model cards identify OpenMDW 1.1. Meta uses the custom Muse Model License with additional commercial terms, and its Hugging Face download requires an approved access request.[6][4][5][1][2]