04 min reading in—AI Hardware

Local LLMs on Strix Halo 128GB Shared Ram: My Tests

Strix Halo 128GB RAM, 100% local LLM agents, my tests

—

Published at Feb 22, 2026

Local LLMs on Strix Halo 128GB Shared Ram: My Tests

Local LLMs on Strix Halo; AMD Ryzen™ AI MAX+ 395 with 128GB Shared Ram: My Tests

LLMs are sometimes stupid, sometimes impressive, often unpredictable and almost never private. Bigger models are generally hosted in clouds and proprietary, sometimes you can limit access or censor data presented to model providers but almost never you can restrict access fully or be really sure how and when your data is being used.

These are just some of the reasons why you might want to run models locally. But hardware especially graphics cards that can run these models are very pricey and the vram requirement for the more 'mid-range' models just exceeds what is possible with local models.

For some time Apple had an alternative to expensive graphics cards with their m-chip lineup allowing shared VRAM with integrated graphics and CPU, but the prices for higher RAM Mac configurations are still ridiculously high. But now there is a 'cheaper' option available for consumers. For a tb-software project I tested the highest available 128GB RAM configuration of a mini PC with the new strix-halo chip to test it and its ability for running local LLMs.

This article reports our early findings and summarizes some configuration steps and the measurement setup I used. To this point I've tested 5 LLMs and more tests are coming. I also investigated the use and usability of local agents with OpenCode in a companion post: Opencode; Usability with Local LLMs on iGPU w 128GB VRAM.

Setting up ollama with vulkan

First I needed to allocate VRAM to the integrated graphics card, I have not played with the 'dynamic' setting here yet. ( I have now tried extending VRAM to 94GB more below ) For these tests I have chosen to set the VRAM to 65536 MiB so there is about 67 GB RAM still available to the CPU. We can also test influence on load speed and inference of varying the available amount of VRAM in the future.

Tests were run on NixOS and to access all the libraries required to use the Ollama Vulkan backend I had to install some additional packages and libraries, specifically I added vulkan-loader to the NixOS system libraries and ran Ollama with OLLAMA_LLM_LIBRARY=vulkan OLLAMA_VULKAN=1 ollama serve. I used amdgpu_top to monitor the VRAM and graphics usage ( It feels like there are still some performance gains to be made, also VRAM is not fully utilized, especially by the smaller models ).

The Test Setup

To test the LLMs for now I just wanted to know basic results for load speed and completion speeds. I've created a simple fork of ollama benchmark ( here my fork without telemetry ) to measure the results based on a running Ollama backend.

If you had the exact same hardware, system and configurations, tests could be replicated via:

python3 -m venv venv
source venv/bin/activate
pip install git+https://github.com/tbscode/tims-ollama-bench-fork.git
tims_llm_benchmark run --custombenchmark=<your-model-config>.yml

With models file:

file_name: "<config-name>.yml"
version: 2.0.custom
models:
  - model: "<model-name>"

And run them all via

tims_llm_benchmark run --custombenchmark=model-configs/qwen3-coder-next.yaml
tims_llm_benchmark run --custombenchmark=model-configs/nemotron-3-nano.yaml
tims_llm_benchmark run --custombenchmark=model-configs/glm-4.7-flash.yaml
tims_llm_benchmark run --custombenchmark=model-configs/gpt-oss-120b.yaml
tims_llm_benchmark run --custombenchmark=model-configs/qwen-3.5-122b.yaml

Test Results

Based on this benchmark and personal prompts and agents tests I've created this table to summarize the results. There are many more tests to be done and likely also performance improvements through improved configuration, these are just my initial findings and a preliminary review. But I can definitely say that there are functional local models out there that can definitely complete simple tasks locally on this hardware!

Name	Size (GB)	Quality	Load Speed	Tokens / sec
qwen3-coder-next	61GB	👾👾👾👾👾	Fast	`35.094`
nemotron-3-nano	34GB	👾👾👾👾	Very Fast	`63.566`
glm-4.7-flash	40GB	👾👾👾👾	Slow	`50.098`
gpt-oss:120b	70GB	👾👾👾👾	Very Slow	`31.532`
qwen-3.5:122b	81GB	👾👾👾👾	Very Slow	`19.158`
qwen-3.5:9b	6.6GB	👾👾👾👾	Very Slow	`29.52`

All these models are usable for smaller and simpler local agentic and coding tasks, involving basic files manipulations and tool calls e.g.: via open-code. I was especially impressed by how well qwen3-coder-next performed, and also nemotron-3-nano that was incredibly fast and smart for such a small thinking model.