Contents
Two of the biggest names in AI hardware and software released free, downloadable models this week, and both made the same pitch: you can run this on your own machine. Meta's new flagship needs two Mac Studios wired together. Nvidia's smallest one fits in 4GB. In between sits a new category of Windows laptop that ships today.
I spent the week sorting out what each release actually needs, in dollars and gigabytes, so you do not have to.
What landed this week
On 15 October, Meta released Muse Spark Open, an 816 billion parameter model (42 billion active per token) with a 1 million token context window, under a licence that lets most people use it for free. Meta calls it the most capable US open weight model to date, and says it is the first model of its size that runs on a desk, using two Apple M5 Ultra Mac Studios.
The same day, Nvidia released Nemotron 3.5 Super VL, a 120 billion parameter model (12 billion active) that reads text, images and short videos, plus two smaller Nemotron 3.5 Bolt models: a 74 billion parameter version compressed to fit in 24GB, and a 4GB version arriving on 23 October that is built for voice assistants and in-app characters.
And today, 16 October, the first RTX Spark laptops start shipping. These are Windows machines built around Nvidia's N1X chip, an Arm processor with a Blackwell GPU and up to 128GB of memory shared between the two. Nvidia built the Bolt models for them.
Three releases, one message: the biggest labs now expect you to run serious AI without the cloud. The question is what that costs.
Open weights in one paragraph
An open weight model is one you can download as a file and run yourself, no account and no subscription. The weights are the numbers the model learned during training, and they are the whole model. I explained where these files live in my Hugging Face explainer, and both of this week's releases are there. What open weight does not mean is "small". Muse Spark Open's full precision weights are about 1.6TB. Downloading is the easy part.
The hardware ladder: what each model actually needs
Every model this week is a mixture of experts design, which means only a slice of it works on each word. That keeps them fast, but every parameter still has to sit in memory. If you read my mixture of experts explainer, this is the same trade: memory decides whether a model runs at all, and the active parameter count decides how fast it feels.
Here is the ladder, bottom to top.
4GB: Nemotron 3.5 Bolt E4B
Nvidia's smallest model, arriving 23 October. It is aimed at fast conversational agents, game characters and local voice assistants rather than at writing your essays. Any gaming laptop from the last few years has room for it. If you have run a small model on your phone using the method in my offline AI guide, this is the same tier of hardware.
24GB: Nemotron 3.5 Bolt 74B-A9B
This is the one most readers can realistically use. It is a 74 billion parameter model with 9 billion active, shipped with native 2-bit compression so the whole thing fits in 24GB and, in Nvidia's words, runs alongside other applications. It works in Ollama, LM Studio and llama.cpp from day one. Nvidia says it ranks first among open models in its class on PinchBench and second on BrowseComp, which are agent benchmarks about browsing and completing tasks rather than chat.
What you need: a desktop graphics card with 24GB of memory, or an RTX Spark laptop with 32GB or more. On a 24GB card you will be close to the ceiling once you add a long conversation, so expect to keep the context short. If 2-bit sounds alarming after my quantization post, note that this is a model trained to live at 2-bit, not a 16-bit model crushed after the fact. That is a different situation from the 1-bit downloads I warned about.
70GB: Nemotron 3.5 Super VL
The 120 billion parameter model that the Bolt models were distilled from. LM Studio's Q4_K_M download is 70.10GB and LM Studio says you need at least 70GB of RAM before you add images or a long context. Nvidia's own 4-bit NVFP4 version is built for one DGX Spark, which has 128GB of unified memory.
What you need: a 128GB machine. That is a top-spec RTX Spark laptop, a DGX Spark, or a Mac Studio with the M5 Max chip and 128GB. This is the tier where a model can read a PDF, look at a screenshot and work through a multi-step task entirely on your desk.
300GB and up: Muse Spark Open
Meta's 4-bit NVFP4 version is about 530GB. The smallest community version, Unsloth's roughly 2-bit GGUF, needs about 300GB of combined RAM and graphics memory according to Meta's own developer page. Nothing you would call a consumer PC gets there. A 512GB Mac Studio does, which is why Meta's headline setup is two of them.
The two Mac Studios math
Meta's recipe is two M5 Ultra Mac Studios with 512GB of unified memory each, connected over Thunderbolt 5, running Apple's MLX framework. Each machine holds half the model and computes half of every token, and only small intermediate results cross the cable. Meta also lists a four machine option with 256GB each. Either way you are buying a terabyte of unified memory.
Here is what Apple has published. The M5 Ultra Mac Studio starts at $5,499 with 96GB. The 256GB configuration adds $4,000, so $9,499 before storage. Apple has not posted the price of the 512GB configuration yet, and that build ships later this month. Meta's four machine route is therefore the only one you can price today: $37,996 before you add a single terabyte of storage or a Thunderbolt 5 cable.
The two machine route will be cheaper, because Apple's memory upgrades scale per machine, but it will not be cheap. Realistically you are looking at a bill north of $20,000 for a model you can also rent by the token from Baseten, Fireworks, Groq, OpenRouter or Together AI, all of which Meta lists as hosting it from launch.
So who is this for? Developers who want a frontier-class coding model that never leaves the building, and small studios with a confidentiality clause. Meta's own demo was fixing bugs and shipping features in an iOS app with its Muse Code terminal, entirely on device. That is the use case.
One more thing. Meta's "most capable US open weight model" line is carefully worded. On Meta's own comparison table, Muse Spark Open scores 39 on the Artificial Analysis Intelligence Index, behind GLM 5.3 at 45, Kimi K3 at 44 and Qwen3.8 at 40, all Chinese releases. It leads them on several agent and vision benchmarks, but it is not the best open model in the world by Meta's own numbers.
The RTX Spark laptop twist
The reason Nvidia's Bolt models exist is a laptop launch. RTX Spark machines from Microsoft, ASUS, HP, Lenovo, Dell and MSI start shipping today at $2,599.99 for a 24GB Surface Laptop Ultra or ASUS ProArt P14, and run up to $6,999.99 for a 128GB ASUS ProArt P16 with 2TB of storage. HP's 128GB OmniBook X 14 is $5,109.
The idea borrowed from Apple is unified memory: one pool shared by the CPU and GPU, so a 128GB laptop can hold a 70GB model that no gaming laptop with a separate 16GB graphics card could touch. Nvidia says 128GB can hold a 120 billion parameter model, and Super VL is exactly that.
The catch is bandwidth. Nvidia's own documentation puts the N1X at about 300GB/s of memory bandwidth. The M5 Ultra is 1.2TB/s, four times higher, and a desktop graphics card is higher again. Bandwidth decides how fast words appear once a model is loaded, the same reason my RAM guide keeps saying that capacity and speed are different purchases. A 128GB RTX Spark laptop can hold Super VL. It will not run it quickly. The 24GB Bolt model, with only 9 billion parameters active per token, is the one that will feel snappy here, and that is not a coincidence.
Also remember this is Windows on Arm. If you depend on a niche piece of software, check it runs before you spend $3,000.
What I ran, and how it went
I loaded Bolt 74B-A9B into LM Studio on my desktop, which has a 24GB graphics card, and it fit with about 2GB to spare at an 8,000 token context. Generation ran at roughly 38 tokens per second, fast enough that it never felt like waiting. I gave it the three tasks I use on every local model: summarise a 12 page PDF, write a Python script that renames files by date, and plan a three day trip on a budget. It handled all three and was noticeably better at the planning task than the 30 billion parameter models I usually run. Push the context past 12,000 tokens and it spilled into system memory and slowed to a crawl, so the 24GB figure is honest but tight.
I could not run Muse Spark Open at home, so I tried it through OpenRouter with a coding task I know well: fix a flaky date parsing function and add tests. It got there in one pass. That says nothing about how it feels on two Mac Studios, but it says the model is real and competent.
The licence fine print
Nvidia's models ship under the OpenMDW licence, which is about as permissive as these get. Meta's Muse Model License is more particular. Anything you build with it has to display "Built with Muse Spark Open" somewhere in the product or its documentation. Companies that sell access to the model as a service and pass $20 million in revenue over any 12 month period have to sign a separate agreement with Meta. Internal use, research and using it as a component of your own product are exempt. The model is also not intended for anyone under 18. For a hobbyist, none of this matters. For a small business shipping a product, read it.
Which one should you download?
If you own a gaming PC with a 24GB card, or you are buying an RTX Spark laptop, download Bolt 74B-A9B today. It is the first big-lab model built specifically to be good at 24GB, and it supports the agent-style tasks (browsing, tool use, multi-step jobs) that small local models have been bad at.
If you have 128GB of unified memory, or you are about to spend on it, Super VL is the model that justifies the spend. Vision input, a million token context and the same family as Bolt, so your prompts carry over.
If you were hoping to run Muse Spark Open yourself, wait. Apple has not priced the 512GB Mac Studio, Meta's own numbers show it trailing the Chinese open models on general intelligence, and you can rent it per token today for a fraction of the hardware bill. Revisit it when 2-bit community versions settle and the Mac Studio price is real.
If you only ever use ChatGPT, Claude or Gemini in a browser, none of this changes your week. It does change what those companies can charge, because every open model this good makes the subscriptions compete harder. That part is good news for everyone.
The bottom line
This week's real news is not that Meta built a model that needs two Mac Studios. It is that Nvidia built one that is good at 24GB and built laptops to carry it. If you already own the hardware, grab Bolt 74B-A9B this weekend and see how far a local agent gets you. If you do not, do not buy anything yet. Prices on both the Mac and the RTX Spark side are moving, and the models you can run are getting better faster than the hardware is getting cheaper.
No comments:
Post a Comment
Note: Only a member of this blog may post a comment.