cd /news/large-language-models/so-what-local-inference-models-are-w… · home topics large-language-models article
[ARTICLE · art-113245] src=forum.level1techs.com ↗ pub= topic=large-language-models verified=true sentiment=· neutral

So, what local inference models are we using?

A user reports running Gemma4:26b on a GPU at 2000+ tokens/s pre-fill and 80+ tokens/s evaluation, and Laguna S 2.1 118B on a CPU server at 13-14 tokens/s, expressing frustration at the lack of models between 30B and 120B parameters that fit in 64GB VRAM. The user calls for a modern MoE model in the 55B-75B range to optimize their hardware.

read2 min views1 publishedAug 27, 2026

So, I am not ready to trust Chinese models yet, and have thus been keeping things western.

I have Gemma4:26b running super fast on my GPU (2000+ tokens/s pre-fill, and 80+ tokens/s evaluation output. I generally use it for web assisted research and problem solving.

Then I have a secondary model (Laguna S 2.1 118B) running on my CPU on my server. Normally this would be extremely frustrating, but with 8 channels of DDR4, I get a total memory bandwidth pretty close to the likes of an Nvidia Spark or a Strix 395+.

CPU’s can’t keep up with the matrix math of even a low end GPU, so pre-fill can be a little slow, but eval output actually isn’t bad on the CPU, with some 13-14 tokens/s

I generally use the heavier Laguna logic model to do things like code review or general purpose reasoning where Gemma4 can’t keep up, and I have spare time to run it in the background.

I’d be curious to hear what others may be using.

I am looking for better models for my use case, but right now I feel stuck.

I am finding that the modern 30B-class models - while excellent for their size still don’t perform anywhere near as well as a good 120B-class model like Laguna S 2.1

My main gripe is that there is very little in-between the 30B-class and the 120B-class. The 30B is kind of small, using only ~27GB of my 64GB GPU RAM with 8bit quantization.120B-class is too large to fit in my 64GB of VRAM even at 4 bit quantization, and thus I can’t benefit from GPU acceleration.

I would really love a modern MoE model in the 55B-75B range which I could run at 4-6bit quantization and make the most out of my VRAM with, but there just isnt much there right now.

I did briefly run the old llama 3.3B 70B model, but it is actually out-performed logic wise by the newer generation of 30B-class models. I just can’t help but wonder how good this new generation of models could be if they could have weights in that size class.

I’m hoping someone comes up with a model like that in the not too distant future.

── more in #large-language-models 4 stories · sorted by recency
── more on @gemma4:26b 3 stories trending now
sponsored brought to you by zahid.host 4,200+ EU-deployed projects
reading about agents? ship yours in a single git push.

Run your AI side-project on zahid.host

EU-based hosting, git-push deploys, automatic HTTPS, no cold starts. Free tier with a custom domain — perfect for shipping the agent you just read about.

$git push zahid main
Live at https://your-agent.zahid.host
Get free account → Pricing
from €0/mo · no card required
LIVE [news/so-what-local-infere…] indexed:0 read:2min 2026-08-27 ·