Two frontier-scale open-weight models landed a day apart, and both are mixture-of-experts designs: memory follows the total parameter count, while speed follows the few parameters active for each token. The second memory bill, the cache that grows with context, is where the two differ by roughly a factor of seven.
Meta’s Muse Spark Open has 816 billion parameters and runs on a 512GB Mac at 2-bit
Muse Spark Open has about 816 billion parameters, around 42 billion of them active per token, and a context window of up to one million tokens, according to Hugging Face’s launch post for the model. It takes text, images, audio and video and writes text. All 256 experts in each layer stay in memory while eight run per token, which puts the BF16 checkpoint at about 1.5 TB. The NVFP4 checkpoint is the 4-bit option for Blackwell GPUs: the post runs it on a single B200 node in transformers and on four GB300 GPUs in the SGLang reference setup, and its Inference Endpoints recipe uses eight RTX Pro 6000 cards with a context length of 131072 tokens.
The numbers a home lab can use come from Apple hardware. On an M3 Ultra with 512GB, an imatrix-calibrated 2-bit llama.cpp quantization uses 304 GB and generates 18 tokens a second. An oMLX 3-bit quantization takes about 344 GB and generates 21 tokens a second after a 1024-token prompt, with prompt processing at 225 tokens a second.
Context is where that Mac runs out of room. The post puts the BF16 cache at about 250 KB per token, or about 266 GB for a full million-token window. Added to the 304 GB of 2-bit weights, one full-length sequence would need about 570 GB, which is more than the machine has unless the cache is quantized too. That sum is our arithmetic from the post’s figures.
Reflection’s Beam is a 501B Apache 2.0 model that needs about 280 GB at 4-bit
Beam is the first model from Reflection, a new US lab, and Hugging Face’s post describes it as a text-only model for coding and agent workflows with 501 billion parameters, 23 billion active per token, a one-million-token window and an Apache 2.0 license. The BF16 checkpoint is about 1 TB. NVFP4 shrinks it to roughly 280 GB but needs Blackwell, so FP8 checkpoints cover Hopper and Ampere cards. Two B200s with 180 GB each can hold the weights and little else, so the post recommends four, with a full 8xB200 node, two B300s at 288 GB each or 8xRTX Pro 6000 as alternatives.
The cache is the smaller bill here. Only 9 of Beam’s 52 layers attend to the whole sequence, and the other 43 keep a 512-token sliding window, so the cache grows by about 36 KiB per token and reaches 36.08 GiB at one million tokens in BF16. That is roughly a seventh of Muse Spark Open’s 266 GB. Each layer caches the same 2,048 values per token in both models, so the gap comes from Muse Spark Open keeping 62 full-length layers against Beam’s 9.
Reflection’s benchmark table, reproduced in the post, shows little cost from quantization. SWEBench Verified reads 81.6% in BF16, 80.2% in FP8 and 80.7% in NVFP4. Terminal Bench v2.1 reads 80.1%, 78.9% and 78.2%.
NVIDIA’s Nemotron 3.5 Bolt 74B-A9B is built to fit in 24GB
At the other end, NVIDIA’s announcement presents Nemotron 3.5 Bolt 74B-A9B 24GB as a model for local agent harnesses on RTX/DGX Spark. It uses native 2-bit quantization to fit in 24GB of memory and leave room for other applications, and it is available now. NVIDIA ranks it first among open models in its class on PinchBench and second on BrowseComp. Bolt E4B 4GB, for GeForce RTX, Jetson and RTX/DGX Spark, follows on October 23. NVIDIA released the pair with Nemotron 3.5 Super VL, a 120-billion-parameter model aimed at servers.