Same, getting ~3 trillion parameters in consumer hardware is rough.

If Nvidia has any foresight they’ll see the writing on the wall and start getting higher memory Spark style SMB inference machines, few people in the long run are going to pay retail API token costs,

Especially giving the lack of trust organizations rightfully have in big AI companies

Exactly, having the mustard toddler pull fable because of, who knows? That burned trust in the west’s AI companies.

And there is also the question of what they’re doing with the data you send into the models.

The other problem is bandwidth. Even if you can run Kimi K3 locally, AI usage patterns tend to be very bursty. Having the scalability of cloud compute to run multiple agents and sessions at once makes it much more usable.

My current workflow is passing a human written spec to an agent to implement with strict coding guidelines, architectural decisions, etc. The agent isn’t making any decisions about the abstractions to use, it’s just creating the objects and test suites. So I don’t mind the slower bandwidth because I’m running the heavy agentic lifting over night with no need for human supervision.

But I fully appreciate that my workflow isn’t the norm. In fact my workflow it’s the exact opposite the AI grifters like Sam Altman are selling because it still involves a human with knowledge of the systems making different decisions.

If you’re running stuff overnight on a homelab, you can get away with using very large models as agents in CPU RAM.

Another option is a hybrid approach: farm “dumb” privacy-insensitive calls to an API you can subscribe to for like $40 a year, and let your local LLM do the orchestrating. This is what I do: I run MiMo locally, but also farm calls out to the exact same LLM in the cloud when appropriate… And I don’t mind one bit. I’m happy to tip them for open sourcing it in the first place.

How much RAM do you have?

I can run MiMo 2.5 at about 9 tokens/sec, on 128GB RAM, a 7800 and a 3090 in an SFF rig. That’s a fantastic 310B model. I’m requantizing it right now, to see if I can speed it up with Dflash.

Still fantastic models can be run on 64GB or 32GB CPU RAM, as long as you have some GPU. We’re awash in sparse models these days.

I’ve got 128GB RAM + 24GB VRAM on a 4090. I’ve managed to get a 400B parameter model running on a single board computer with 64GB RAM by using MMAP. But I want to run Kimi K3 locally so I would need a lot more RAM / bandwidth

Oh that’s perfect!

You can’t realistically run Kimi (as it’s a 1T+ model), but you’re set. There’s a glut of excellent 120B-300B models for you to choose from.

I’m quantizing MiMo 2.5 specifically to cram into 24GB/128GB as I type this. It’s tight, and will use up most of your memory, but it’s fantastically smart and plenty fast. The quantization won’t finish cooking until tomorrow, but I’ll upload it to huggingface then.

But there are others existing quants would fit, like Deepseek Flash IQ3_S: huggingface.co/unsloth/…/UD-IQ3_S

Or… well, any of these! huggingface.co/models?num_parameters=min%3A128B%2…

The key for you is to run the ik_llama.cpp fork: github.com/ikawrakow/ik_llama.cpp/

It’s specifically optimized for hybrid (CPU + GPU) inference on Nvidia desktops; you’ll get MUCH faster speeds than mainline llama.cpp or anything based on it. It also supports some more exotic quantization type; as an example, I’m quantizing MiMo 2.5 as a hybrid quant, with the dense layers at IQ6K/Q8_0 and the sparse experts as an IQ3_KT “trellis” quant type. This should yield a higher fidelity quantization than a typical Q3 GGUF while taking less RAM, at the cost of taking forever to quantize and a slight speed hit.

Some other quantization types (like the KS or R4 types) are specifically configured to be fast on CPU.

There are specialized “quant cookers” that make GGUFs specifically for ik_llama.cpp, like:

huggingface.co/ubergarm

huggingface.co/AesSedai

huggingface.co/sigargv/Laguna-M.1-GGUF

huggingface.co/models?other=ik_llama.cpp&sort=mod…

Mine will be here once I upload it:

huggingface.co/Downtown-Case

For anyone else reading this: none of this is applicable to you if you have an AMD/Intel GPU, or an older Nvidia GPU, or less than a certain amount of RAM, or a non AVX2 CPU or… well, there’s all sorts of caveats.

The optimal runtime is different for everyone. As an example, exllamav3 is WAY better than llama.cpp on modern Nvidia GPUs until you get above a certain amount of CPU RAM; then using all that RAM for hybrid inference makes more sense. On AMD, different backends work better on different GPUs, and… well, you get the point. Basically all LLM running advice is irrelevant without specifics of your hardware, even this post will be obsolete in a month.