What's also new here, is VRAM-context size trade-off: for 25% of it's attention network, they use the regular KV cache for global coherency, but for 75% they use a new KV cache with linear(!!!!) memory-token-context size expansion! which means, eg ~100K token -> 1.5gb VRAM use -meaning for the first time you can do extremely long conversations / document processing with eg a 3060.
Strong, strong recommend.
I like the idea of finding practical uses for it, but so far haven't managed to be creative enough. I'm so accustomed to using these things for programming.
If you want a general knowledge model for answering questions or a coding agent, nothing you can run on your MacBook will come close to the frontier models. It's going to be frustrating if you try to use local models that way. But there are a lot of useful applications for local-sized models when it comes to interpreting and transforming unstructured data.
Azure Doc Intelligence charges $1.50 for 1000 pages. Was that an annual/recurring license?
Would you mind sharing your OCR model? I'm using Azure for now, as I want to focus on building the functionality first, but would later opt for a local model.
Try just GLM-OCR if you want to get started quickly. It has good layout quality, good text recognition quality, and they actually tested setup on Apple Silicon laptops. It works easily out-of-the-box without the long yak shaving setups I encountered with some other models I tried. Chandra is even more accurate on text but its layout bounding boxes are worse and it runs very slowly unless you can set up batched inference with vLLM on CUDA. (I tried to get batching to run with vllm-mlx so it could work entirely on macOS, but a day spent shaving the yak with Claude Opus's help went nowhere.)
If you just want to transcribe documents, you can also try end-to-end models like olmOCR 2. I need pipeline models that expose inner details of document layout because I need to segment and restructure page contents for further processing. The end-to-end models just "magically" turn page scans into complete Markdown or HTML documents, which is more convenient for some uses but not mine.
Trying to run LLM models somehow makes 128 GiB of memory feel incredibly tight. I'm frequently getting OOMs when I'm running models that are pushing the limits of what this can fit, I need to leave more memory free for system memory than I was expecting. I was expecting to be able to run models of up to ~100 GiB quantized, leaving 28 GiB for system memory, but it turns out I need to leave more room for context and overhead. ~80 GiB quantized seems like a better max limit when trying not running on a headless system so I'm running a desktop environment, browser, IDE, compilers, etc in addition to the model.
And memory bandwidth limitations for running the models is real! 10B active parameters at 4-6 bit quants feels usable but slow, much more than that and it really starts to feel sluggish.
So this can fit models like Qwen3.5-122B-A10B but it's not the speediest and I had to use a smaller quant than expected. Qwen3-Coder-Next (80B/3B active) feels quite on speed, though not quite as smart. Still trying out models, Nemotron-3-Super-120B-A12B just came out, but looks like it'll be a bit slower than Qwen3.5 while not offering up any more performance, though I do really like that they have been transparent in releasing most of its training data.
There's also the Ryzen AI Max+ 395 that has 128GB unified in laptop form factor.
Only Apple has the unique dynamic allocation though.
I'm not sure what exactly you're referring to with "Only Apple has the unique dynamic allocation though." On Strix Halo you set the fixed VRAM size to 512 MB in the BIOS, and you set a few Linux kernel params that enable dynamic allocation to whatever limit you want (I'm using 110 GB max at the moment). LLMs can use up to that much when loaded, but it's shared fully dynamically with regular RAM and is instantly available for regular system use when you unload the LLM.
I configured/disabled RGB lighting in Windows before wiping and the settings carried over to Linux. On Arch, install & enable power-profiles-daemon and you can switch between quiet/balanced/performance fan & TDP profiles. It uses the same profiles & fan curves as the options in Asus's Windows software. KDE has native integration for this in the GUI in the battery menu. You don't need to install asus-linux or rog-control-center.
For local AI: set VRAM size to 512 MB in the BIOS, add these kernel params:
ttm.pages_limit=31457280 ttm.page_pool_size=31457280 amd_iommu=off
Pages are 4 KiB each, so 120 GiB = 120 x 1024^3 / 4096 = 31457280
To check that it worked: sudo dmesg | grep "amdgpu.*memory" will report two values. VRAM is what's set in BIOS (minimum static allocation). GTT is the maximum dynamic quota. The default is 48 GB of GTT. So if you're running small models you actually don't even need to do anything, it'll just work out of the box.
LM Studio worked out of the box with no setup, just download the appimage and run it. For Ollama you just `pacman -S ollama-rocm` and `systemctl enable --now ollama`, then it works. I recently got ComfyUI set up to run image gen & 3d gen models and that was also very easy, took <10 minutes.
I can't believe this machine is still going for $2,800 with 128 GB. It's an incredible value.
What do you mean? On Linux I can dynamically allocate memory between CPU and GPU. Just have to set a few kernel parameters to set the max allowable allocation to the GPU, and set the BIOS to the minimum amount of dedicated graphics memory.
Apple has none of this.
Setting the kernel params is a one-time initial setup thing. You have 128 GB of RAM, set it to 120 or whatever as the max VRAM. The LLM will use as much as it needs and the rest of the system will use as much it needs. Fully dynamic with real-time allocation of resources. Honestly I literally haven't even thought of it after setting those kernel args a while ago.
So: "options ttm.pages_limit=31457280 ttm.page_pool_size=31457280", reboot, and that's literally all you have to do.
Oh and even that is only needed because the AMD driver defaults it to something like 35-48 GB max VRAM allocation. It is fully dynamic out of the box, you're only configuring the max VRAM quota with those params. I'm not sure why they choice that number for the default.
On Windows, I think you're right, it's max 96 GiB to the GPU and it requires a reboot to change it.
An OpenRunPod with decent usage might encourage more non-leading labs to dump foundation models into the commons. We just need infra to run it. Distilling them down to desktop is a fool's errand. They're meant to run on DC compute.
I'm fine with running everything in the cloud as long as we own the software infra and the weights.
This is conceivably the only way we could catch up to Claude Code is to have the Chinese start releasing their best coding models and for them to get significant traction with companies calling out to hosted versions. Otherwise, we're going to be stuck in a take off scenario with no bridge.