Strata

Strata

Docker app from MadeOfCard's Repository

Overview

Strata runs the Qwen3.8-Flash-Next mixture-of-experts model on one NVIDIA GPU plus system RAM, and serves an OpenAI- and Anthropic-compatible API, a chat page and a live monitor. Requirements: the Nvidia Driver plugin on driver 580 or newer; an RTX 20/30/40/50 card with 12 GB of VRAM or more; an AVX2 CPU; 48-62 GB of free RAM depending on the model; 70-90 GB of disk on an SSD pool. First start: downloads the model (about 70 GB for IQ2_XS) and sets it up, which can take a few hours. Watch the log; it resumes if it stops. Later starts go straight to serving, and loading the model into RAM takes 1-3 minutes. Changing settings later: switching Model or Model family needs nothing extra. Changing context, vision, KV cache, API key or low-RAM mode for a model already set up needs Reinstall = 1 for one start (nothing is downloaded again), then set it back to 0. No API key by default: anyone who can reach the port can use the model. Keep it LAN-only, or set an API key. This is an unofficial image built from Strata's own Dockerfile. Report template or image problems at the Support link, not to the Strata project.

Readme

View on GitHub

Strata for Unraid

An Unraid Community Applications template and container image for Strata, which runs the Qwen3.8-Flash-Next mixture-of-experts model on one NVIDIA GPU plus system RAM. It serves an OpenAI- and Anthropic-compatible API, a chat page and a live monitor.

This is an unofficial image. Report template or image problems in this repository's issues, not to the Strata project.

Install

Search for Strata in the Apps tab. Before installing you need:

  • the Nvidia Driver plugin on driver 580 or newer
  • an NVIDIA RTX 20/30/40/50 card with 12 GB of VRAM or more
  • an x86-64 CPU with AVX2
  • 48-62 GB of free RAM, depending on the model
  • 70-90 GB of disk on an SSD pool, mapped as a direct pool path (/mnt/cache/...), not /mnt/user

The first start downloads the model and sets it up, which can take a few hours. Later starts go straight to serving. The template's Overview explains each setting.

The server has no API key by default: anyone who can reach the port can use the model. Keep it on your LAN, or set the API key.

Updating to a new Strata release

Change the tag in build/STRATA_VERSION and merge it to main. If the patch no longer applies, upstream changed its build step: redo the edit in the new Dockerfile and regenerate the patch with git diff.

Credits and licence

Strata is by Niko1221 and the Strata contributors, under the MIT licence. The icon is Strata's own. This repository is MIT licensed; see LICENSE.

Requirements

Nvidia Driver plugin on driver 580 or newer. Do not pin CPUs: the engine sets its own thread affinity and breaks inside a cpuset. Do not set a memory limit unless Low-RAM mode is on.

Categories

AI

Related apps

Explore more like this

Explore all
AI

Links

Projectgithub.comSupportgithub.comRegistrygithub.comTemplateraw.githubusercontent.com

Details

Repository
ghcr.io/williamkunz/strata:latest
Registry
https://github.com/williamkunz/strata-docker/pkgs/container/strata
Last Updated2026-10-15
First Seen2026-10-15

Runtime arguments

Web UI
http://[IP]:[PORT:8080]/
Network
bridge
Shell
bash
Privileged
false
Extra Params
--runtime=nvidia --ulimit memlock=-1 --stop-timeout 60

Template configuration

WebUI / API portPorttcp

Chat page, monitor and the /v1 API. No API key is set by default, so keep this LAN-only.

Target
8080
Default
8080
Value
8080
DataPathrw

The model, the prepared pack, the MTP layer and the install config: 70-90 GB depending on the model. Use a direct pool path (/mnt/cache/... or /mnt/POOLNAME/...), not /mnt/user: the engine reads experts from these files while it runs.

Target
/data
Default
/mnt/cache/appdata/strata
Value
/mnt/cache/appdata/strata
GPUVariable

Your card's UUID (GPU-xxxxxxxx-...) from the Nvidia Driver plugin page or nvidia-smi -L. Name one card; 'all' on a two-card server makes setup split the model across both.

Target
NVIDIA_VISIBLE_DEVICES
ModelVariable

IQ2_XS (recommended): 68 GB download, 48 GB RAM. Q2_0: fastest, 48 GB RAM (qwen only). IQ3_XXS: better, 60 GB RAM. IQ3_S: best quality, 62 GB RAM. IQ1_M: coder family only, 32 GB RAM. UD-Q4_K_XL: unsloth family only.

Target
MODEL
Default
IQ2_XS|Q2_0|IQ3_XXS|IQ3_S|IQ1_M|UD-Q4_K_XL
Value
IQ2_XS
Model familyVariable

qwen (default); swift (thinks shorter); coder (for code, fewer experts); unsloth and nvfp4 need their own models (see Model).

Target
FAMILY
Default
qwen|swift|coder|unsloth|nvfp4
Value
qwen
Context lengthVariable

Tokens per conversation. Longer costs RAM; setup caps it if the RAM is short. Changing it for a model already set up needs Reinstall = 1.

Target
CONTEXT
Default
32768|8192|65536|131072|204800|262144|393216|524288
Value
32768
VisionVariable

Image input: yes runs the image encoder on the GPU, cpu runs it on the CPU. Changing it for a model already set up needs Reinstall = 1.

Target
VISION
Default
no|yes|cpu
Value
no
API keyVariable

Empty = no key. When set, /v1/* requires it as Authorization: Bearer or x-api-key; /health and the chat page stay open. Setting or changing it for a model already set up needs Reinstall = 1.

Target
API_KEY
ReinstallVariable

1 = run setup again on the next start to apply changed context, vision, KV cache, API key or low-RAM settings. Nothing is downloaded again. Set it back to 0 afterwards.

Target
REINSTALL
Default
0|1
Value
0
KV cache typeVariable

int8 (default); q4_0 halves the KV memory, slightly less precise; k8v4 keeps its KV in VRAM from 64K context up. Needs Reinstall = 1 to change.

Target
KV
Default
int8|q4_0|k8v4
Value
int8
Low-RAM modeVariable

on maps the experts from the pack on disk instead of keeping them in RAM. Required if you give the container a memory limit: setup sees the server's total RAM, not the limit. Needs Reinstall = 1 to change.

Target
LOW_RAM
Default
auto|on|off
Value
auto
NVIDIA driver capabilitiesVariable

Must include utility: setup needs nvidia-smi.

Target
NVIDIA_DRIVER_CAPABILITIES
Default
compute,utility
Value
compute,utility