
Alexander is a writer, traveler, and serial digital entrepreneur. He’s the founder of the Tushi, the Gen X personal growth newsletter Further, and offers small business advice at Unemployable.

Twitter

Linkedin

Facebook

Director and Chief Evangelist

Alexander Hipp


Anil is a content-obsessed marketer who loves driving organic traffic for outstanding companies. He is also the creator of Hack the Content, co-host of The Runner, and owner of several digital commerce sites.

Managing Partner

Anil Vugels


Monika is a writer, musician, and content strategist with a passion for language. When she isn’t helping clients, you can find her writing songs, hiking in a northern forest, or furiously journaling.

Director of Operations

Ann Monika

Mastodon

Github

Programmer / Founder / Student

Tim Schupp

Serving Qwen3.8-27B (INT4 AWQ) and Qwen3-4B (FP16) with vLLM on my two-node k3s cluster, fronted by a LiteLLM OpenAI-compatible router.

vLLM + LiteLLM on AMD Strix Halo

A hands-on experiment building a self-managed at-home AI cluster with k3s, Ollama, and LiteLLM.

Building a private LLM Cluster


<img width="1906" height="917" alt="OpenCode blog cover" src="/static/assets/opencode-blog-cover.png" />

[Opencode](https://opencode.ai/) is a pretty cool and neatly composable terminal based llm coding agent.
For a project at [tb-software](https://timschupp.de/) and also private personal use in general I'm especially interesting in the usage of open-code or similar tools with local llms, hosted on own hardware. See [my article on running and testing local llms on a 128gb strix halo setup here](https://blog.t1m.me/blog/local-llms-on-strix-halo-128gb-shared-ram).
This article will describe my personal setup as well as some hints how to manage it using a nix-flake - this functions more as future reference - than direct instructions. But it may help someone configure open-code for local llm usage.

### Configuring Open-Code

Just create a `opencode.json` like this:

```json
{
  "$schema": "https://opencode.ai/config.json",
  "provider": { ... },
}
```
Then start adding your providers, fistly the local llm one, e.g.: you could use here [ollama]() though there are probaly some better and more performant options out there ( _these I still have to test again / with newer models_ ). But project like localai lammacpp, ... come to mind. Anything that host api compatible with the openai api schema really.

```json
    "ollama": {
      "npm": "@ai-sdk/openai-compatible",
      "name": "ollama-local",
      "options": {
        "baseURL": "http://localhost:11434/v1"
      },
      "models": { ... }
```
e.g.: to test the models from [my latest benchark](https://blog.t1m.me/blog/local-llms-on-strix-halo-128gb-shared-ram) i added. _Of course you will have to pull the models and corretly setup and start the local lmm server first ( check my other post for an example using ollama on vulkan + nix )._

```json
        "qwen3-coder-next": { "name": "qwen3-coder-next" },
        "mdq100/Qwen3-Coder-30B-A3B-Instruct:30b": { "name": "qwen-code-instuct-small"},
        "nemotron-3-nano": { "name": "nemotron-3-nano" },
        "glm-4.7-flash": { "name": "glm-4.7-flash" },
        "gpt-oss:120b": { "name": "gpt-oss:120b" }
```

<img width="1906" height="917" alt="OpenCode post starting prompt example" src="/static/assets/open-code-post-starting-prompt.png" />

### Usability / Review

As already stated in my benchmark post I was impressed by the performance and usability of the small llms, especially on small simple task like file editing, moving things around, manipulating issues, making simple commits and more. *So it is definitely somewhat hopeful that these open models do seems to perform for some use-case*; So I would say with care and good guardrails they can applied for certain tasks; especially when fully private data processing is required.

Also for the full potential of open-code and priming agents to be able to perform certain tasks in certain environments; you should definitely checkout and learn [how open-code mcps](https://opencode.ai/docs/mcp-servers/) and [agent skills](https://opencode.ai/docs/skills/) are set-up. I my experience the small local models perform best with easy; repeatable and easy to verify tasks - especially when given the right context clues and tools to use - but tasks that involve steps hard to statically automate ( e.g.: text based decision; complex semi autonomous task; browser use; search etc... ). Also it is possible to build in restrictive confirmation based tools for oversight.

### External Providers

Also if you want to use other remove non local providers to test oss models, you could add the following. _Note the api keys are taken from environment you don't need to include them in the config_.

#### 1. Deepinfra

```json
    "deepinfra": {
      "npm": "@ai-sdk/openai-compatible",
      "name": "DeepInfra",
      "options": {
        "baseURL": "https://api.deepinfra.com/v1",
        "apiKeyEnv": "DEEPINFRA_API_KEY"
      },
      "models": { ... }
```

#### 2. groq

```json
    "groq": {
      "npm": "@ai-sdk/openai-compatible",
      "name": "Groq",
      "options": {
        "baseURL": "https://api.groq.com/openai/v1",
        "apiKeyEnv": "GROQ_API_KEY"
      },
      "models": { ... }
```

#### a complete config example

```json
{
  "$schema": "https://opencode.ai/config.json",
  "provider": {
    "ollama": {
      "npm": "@ai-sdk/openai-compatible",
      "name": "Ollama (local)",
      "options": {
        "baseURL": "http://localhost:11434/v1"
      },
      "models": {
        "qwen3-coder-next": { "name": "qwen3-coder-next" },
        "mdq100/Qwen3-Coder-30B-A3B-Instruct:30b": { "name": "qwen-code-instuct-small"},
        "nemotron-3-nano": { "name": "nemotron-3-nano" },
        "glm-4.7-flash": { "name": "glm-4.7-flash" },
        "gpt-oss:120b": { "name": "gpt-oss:120b" }
      }
    },
    "deepinfra": {
      "npm": "@ai-sdk/openai-compatible",
      "name": "DeepInfra",
      "options": {
        "baseURL": "https://api.deepinfra.com/v1",
        "apiKeyEnv": "DEEPINFRA_API_KEY"
      },
      "models": {
        "nvidia/Nemotron-3-Nano-30B-A3B": { "name": "nvidia/Nemotron-3-Nano-30B-A3B" },
        "moonshotai/Kimi-K2.5": { "name": "moonshotai/Kimi-K2.5" },
        "Qwen/Qwen3-235B-A22B-Instruct-2507": { "name": "Qwen/Qwen3-235B-A22B-Instruct-2507" },
        "MiniMaxAI/MiniMax-M2.5": { "name": "MiniMaxAI/MiniMax-M2.5" }
      }
    },
    "groq": {
      "npm": "@ai-sdk/openai-compatible",
      "name": "Groq",
      "options": {
        "baseURL": "https://api.groq.com/openai/v1",
        "apiKeyEnv": "GROQ_API_KEY"
      },
      "models": {
        "moonshotai/kimi-k2-instruct": { "name": "moonshotai/kimi-k2-instruct" },
        "qwen/qwen3-32b": { "name": "qwen/qwen3-32b" }
      }
    }
  }
}
```

### Managing a global open-code install via nix

<img width="1906" height="917" alt="OpenCode performing git push" src="/static/assets/open-code-performing-git-push.png" />

To mange my local open-code installation and configuration I've setup a simple nix flake, in a git sub-module that stores and pulls my open-code configuration and makes it globally available on configuration update.

```nix
{
  description = "Install opencode.json as a global OpenCode config";

  inputs = {
    nixpkgs.url = "github:NixOS/nixpkgs/nixos-unstable";
  };

  outputs = { self, nixpkgs }:
    let
      opencodeJsonPath = ./opencode.json;

      commonOptions = { lib, ... }:
        {
          options.opencodeOllamaLocalSetup = {
            enable = lib.mkOption {
              type = lib.types.bool;
              default = false;
              description = "Whether to install the OpenCode config file globally.";
            };
          };
        };

      nixosModule = { config, lib, ... }:
        let
          cfg = config.opencodeOllamaLocalSetup;
          defaultUsers = lib.attrNames (lib.filterAttrs (_: user:
            (user.isNormalUser or false) || ((user.uid or null) != null && user.uid >= 1000)
          ) config.users.users);
          targetUsers = if cfg.users == null then defaultUsers else cfg.users;
          existingUsers = lib.filter (name: lib.hasAttr name config.users.users) targetUsers;
          userRules = lib.concatLists (map (name:
            let
              user = config.users.users.${name};
              homeDir = user.home or "/home/${name}";
              groupName = user.group or name;
            in [
              "d ${homeDir}/.config/opencode 0700 ${name} ${groupName} -"
              "L+ ${homeDir}/.config/opencode/opencode.json - ${name} ${groupName} - ${opencodeJsonPath}"
            ]
          ) existingUsers);
        in
        {
          options.opencodeOllamaLocalSetup.users = lib.mkOption {
            type = lib.types.nullOr (lib.types.listOf lib.types.str);
            default = null;
            description = "Users to receive ~/.config/opencode/opencode.json (null = all normal users).";
          };

          config = lib.mkIf cfg.enable {
            systemd.tmpfiles.rules = userRules;
          };
        };

      homeManagerModule = { config, lib, ... }:
        let
          cfg = config.opencodeOllamaLocalSetup;
        in
        {
          config = lib.mkIf cfg.enable {
            xdg.configFile."opencode/opencode.json".source = opencodeJsonPath;
          };
        }
    in
    {
      nixosModules.default = { ... }: {
        imports = [ commonOptions nixosModule ];
      };
      homeManagerModules.default = { ... }: {
        imports = [ commonOptions homeManagerModule ];
      };
    }
}
```

<img width="1906" height="917" alt="OpenCode blog section edit example" src="/static/assets/opencode-blog-section-edit-example.png" />


Testing and configuring opencode for usage with local llms

Opencode; Usability with Local LLMs on iGPU w 128GB vram: My Tests


## Local LLMs on Strix Halo; [AMD Ryzen™ AI MAX+ 395](https://www.amd.com/en/blogs/2025/amd-ryzen-ai-max-395-processor-breakthrough-ai-.html) 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](https://timschupp.de/) 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](https://blog.t1m.me/blog/opencode-with-local-llms).

### 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.

<img width="1451" height="234" alt="Image" src="/static/assets/strix-halo-vram.png" />

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](https://github.com/aidatatools/ollama-benchmark) ( here [my fork without telemetry](https://github.com/tbscode/tims-ollama-bench-fork) ) 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:

```bash
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:

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

And run them all via

```bash
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
```

<img width="1906" height="917" alt="Image" src="/static/assets/ollama-benchmark.png" />

### 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.

### Testing new qwen-3.5 with increased 94gb vram

Now I've updated the bios to allocate 94GB VRAM to allow larger models to run entirely in VRAM.

<img width="1451" height="234" alt="BIOS VRAM settings for 94GB allocation" src="/static/assets/strix_halo_increase_94gb_vram.png" />

Here's the active usage with 94GB VRAM allocated:

<img width="1906" height="917" alt="94GB VRAM active usage showing model loaded in VRAM" src="/static/assets/strix_halo_94gb_vram_active_usage_strix_halo.png" />

With the increased VRAM, I was able to download and run qwen-3.5:95b entirely in VRAM:

<img width="1451" height="917" alt="Downloading qwen-3.5:95b model with Ollama" src="/static/assets/ollama_downloading_qwen122b.png" />


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

Llm

vLLM + LiteLLM on AMD Strix Halo

Building a private LLM Cluster

Opencode; Usability with Local LLMs on iGPU w 128GB vram: My Tests

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