Gufo
Gufo is an open source local inference engine built and tuned for AMD Strix Halo (Ryzen AI MAX+ 395, Radeon 8060S gfx1151). It serves Qwen3.8, DeepSeek V4 Flash, speech, embedding, image and video models through an OpenAI-compatible API, with speculative decoding, concurrent requests and prompt caching. The script installs the official release build through Nix, using prebuilt ROCm from the Gufo binary cache.
Install
bash -c "$(curl -fsSL https://raw.githubusercontent.com/Heretek-AI/ProxmoxVE/main/ct/gufo.sh)"Run as root in the Proxmox VE shell. Pick Default for the sizing below, or Advanced to choose storage, network and resources yourself.
Notes
- Warning: Gufo targets AMD Strix Halo (gfx1151) only. The container needs /dev/kfd and /dev/dri passed through, which the script sets up when GPU passthrough is enabled.
- The install runs the official Nix installer from nixos.org and asks before it does. Set var_external_installer=yes for unattended installs.
- The first start downloads Qwen3.8 27B Q4_K_XL (about 17 GB). Follow it with: journalctl -fu gufo
- The API requires a key, generated at install. Find it inside the container with: cat ~/gufo.creds. Clients send it as a Bearer token.
- OpenAI-compatible API: http://<IP>:8080/v1
- Change the model or server options in /opt/gufo/gufo.env, then run: systemctl restart gufo
- Check the GPU from inside the container with: gufo diagnose
- If a release is not in the Gufo binary cache yet, installing or updating compiles it locally and takes much longer.
Script details
You might also need
Hermes Agent
LXCAI / Coding & Dev-Tools · debian 13
Hermes Agent is a self-improving AI agent built by Nous Research with persistent memory, autonomous skill creation, computer-use (CUA) support, and multi-platform gateway integration. This enhanced Heretek build includes pre-configured passwordless sudo for the hermes user, native X11/input shared libraries for cua-driver, and Docker Engine with Buildx and Compose v2.
Kyojin
LXCAI / Coding & Dev-Tools · ubuntu 24.04
Kyojin is Yamz Labs' open source inference engine for AMD Strix Halo (gfx1151), built on ExLlamaV3 with speculative decoding, serving their own EXL3 quantised builds of Qwen3.8-Flash-Next, GLM-5.3-Flash and MiMo-V2.6-Flash (95 to 106 GB) on one 128 GB machine through an OpenAI-compatible API. The script builds the latest tagged release with AMD's gfx1151 ROCm wheels, following Kyojin's own container recipe.
GSQHalo.cpp
LXCAI / Coding & Dev-Tools · debian 13
GSQHalo.cpp is a llama.cpp fork tuned for ISTA-DASLab's GSQ-RCO quantizations of Qwen3.8-Flash-Next on AMD Strix Halo (gfx1151), with faster prefill kernels, MTP that fits in memory and a KV cache on SSD. The script installs AMD's ROCm 7.14.1 for gfx1151 in its own prefix, builds the fork from source and serves llama-server's OpenAI-compatible API.
Strata
LXCAI / Coding & Dev-Tools · debian 13
Strata is an open source inference engine for Qwen3.8-Flash-Next and other large MoE models, with a Strix Halo (gfx1151) HIP engine, MTP drafting and an OpenAI-compatible API. The script installs the Strata Terminal, which downloads the signed engine for the GPU with its own HIP runtime, and serves ISTA-DASLab's Qwen3.8-Flash-Next Coder by default.