measuredlocally.comNemotron 3.5 Bolt on an RTX 4090, and Meta's Muse Spark Open
Two big open models came out on Thursday 15 October. NVIDIA's Nemotron 3.5 Bolt fits on my RTX 4090 and tied GLM 5.3 Flash in our eight tests, though its builds look far worse, and Meta's Muse Spark Open is huge.
Plays here, from YouTube
Two big open models came out on Thursday 15 October. The first is NVIDIA's Nemotron 3.5 Bolt 74B, which NVIDIA says is built for running agents on RTX Spark.
NVIDIA says it uses 9 billion of its 74 billion parameters for each token, and that it has squeezed the model into 24GB by storing most of its weights at 2 bits, which it trained the model to cope with.
On my RTX 4090
24GB is exactly the memory of my RTX 4090, and NVIDIA's own llama.cpp instructions only cover graphics cards with 32GB or more. It does fit, with a shorter context than NVIDIA's recipes, 32,768 tokens against their 65,536. The whole model sat on the card, and it wrote at 108.3 to 112.1 tokens a second, depending on the task.
In our eight tests it got exactly the same score as GLM 5.3 Flash, 97.9, in 3.2 minutes on one graphics card. But put their builds side by side and there are some important differences.
GLM drew an SVG fox in a snow forest, and a 3D lighthouse at sunset. Nemotron's fox, on a wide screen, is mostly sky with its head poking out at the bottom, and its lighthouse is a grey tube on a green rock. The score is the same, and the builds show which one I would rather have, which is why it is important to have judges.
Muse Spark Open
The second model is huge. Meta released the weights of Muse Spark Open, which Meta says has 816 billion parameters, 42 billion of them used for each token.
SiliconANGLE reported that Meta ran it on two M5 Ultra Mac Studios, with about a terabyte of memory between them. On World of AI's ranking it comes a couple of points behind Qwen 3.8 27B.
Read more
- NVIDIA's page for Nemotron 3.5 Bolt 74B, with its own recipeshuggingface.co
- NVIDIA's post on the RTX Spark launch and its new modelsblogs.nvidia.com
- Meta's post on Muse Spark Openpytorch.org
- SiliconANGLE's report on Muse Spark Opensiliconangle.com
- World of AI's video with Muse Spark Open in its rankingyoutube.com
The numbers in this story
| Figure | What it is | From | Kind |
|---|---|---|---|
| 24GB | Nemotron 3.5 Bolt's size, with most of its weights at 2 bits | blogs.nvidia.com | NVIDIA's claim |
| 74B, 9B active | Its parameters in all, and those used for each token | blogs.nvidia.com | the maker's claim |
| 108.3 to 112.1 tok/s a task | Its writing speed on my RTX 4090, task by task | my own test | measured by me |
| 97.9 in 3.2 minutes | Its score in our eight tests, the same as GLM 5.3 Flash's, and its time on one card | my own test | measured by me |
| 816B, 42B active | Muse Spark Open's parameters in all, and those used for each token | pytorch.org | the maker's claim |
| about 1TB | The two M5 Ultra Mac Studios Meta ran it on, and their memory between them | siliconangle.com | reported |
If a number on this page is wrong, I would rather know. Tell me in the comments under the video and I will correct it here and pin the correction there. What has been corrected so far is on the corrections page.