Workflows/Run a full local AI studio with Imagination Studio

Video / WORKFLOW

Run a full local AI studio with Imagination Studio

Set up Imagination Studio by keys+Wasserman, a full AI studio for video with sound, images, voice, music, sound effects and editing, from a one-click RunPod template or on a supported high-memory Mac or NVIDIA machine, then make your first clip.

Verified sourceSource checked 11/4/2026

LOCALRENTED GPU

Before you begin

Hardware

What each machine runs with the core install. Local timings and memory are the publisher’s measurements on the named machines; rows marked untested follow from those measurements. The RunPod row is YouRunAI’s own run.

Machine	Imagination Studio 4.3.1 core
Mac Studio M5 Ultra · 256 GB	Tested by the publisher: everything runs. A 5 s 480p FastH3 clip peaks at 28 GiB; 720p 10 s peaks at 32 GiB.
Mac Studio or other Mac · M5 family · 96–192 GB	Should run the default (pictures, FastH3 video with sound, voice, music); 128 GB adds MiniMax H3 Turbo. Untested by the publisher below 256 GB.
Mac · M1 to M4 · 96 GB+	Limited: runs slowly; FastH3 uses the 8-bit MLX engine (about 37–40 GiB peak, roughly six times slower a pass than an M5). Start with 480p, 5 s and 4 passes. Untested on real hardware.
Mac · Apple Silicon · 48–95 GB	Limited, pictures only: install with --image-only (33 GB disk). The installer refuses video under 96 GB.
Mac · under 48 GB	Not supported: the installer stops before downloading. Use the RunPod template instead.
NVIDIA DGX Spark · GB10 · 128 GB	Every feature runs, including LoRA training. A 480p 5 s clip takes about 6.6 s a pass (8 passes) once loaded.
NVIDIA RTX Pro 6000 Blackwell · 96 GB + 64 GB RAM	Core install runs on the card: pictures and FastH3 video with sound. 128 GB of system RAM recommended.
RunPod · RTX PRO 6000 Blackwell · 96 GB VRAM	Recommended: YouRunAI successfully installed the Studio and generated an image and FastH3 video with sound on a RunPod RTX PRO 6000; a fresh deployment from the YouRunAI template completed automatic setup and served the web app (HTTP 200). Not every feature has been re-checked on a fresh template Pod.
NVIDIA RTX 5090 · 32 GB + about 96 GB RAM	Experimental: video runs as an opt-in (STUDIO_ALLOW_SMALL_VRAM=1 and STUDIO_ROOM_MB): about 55 s per 5 s 1344x768 clip at 8 passes, measured on one machine. Requires nvcc 12.9 or newer.
RTX Spark N1X laptop · 128 GB (96 GB GPU carve-out)	Advanced: the whole Studio inside WSL2; about 7.0 s a pass in performance mode, 9.0 s in the default mode; clips match a DGX Spark byte for byte.
Other NVIDIA GPUs (e.g. RTX 4090 · 24 GB)	Not supported: the CUDA engines are built and measured for the GB10 and Blackwell 96 GB cards. Use the RunPod template instead.
Other Windows PCs · 16 GB+	Limited: planning preview only; script import, Slate planning, storyboards and prompt compiling; video generation is not available.

Memory is the limiting factor, not speed. The installers check memory and disk before downloading and stop with the reason when a machine is too small.

LOCAL AI STUDIO / VIDEO WITH SOUND · IMAGES · VOICE · MUSIC · EDITINGSOURCE-LINKED SETUP

Run a full local AI studio with Imagination Studio

Imagination Studio by keys+Wasserman, powered by TensorFold, is a free, Apache-2.0 AI film studio that runs on your own high-memory machine or on a GPU you rent. Built by Keys (drowzeys) with director Sam Wasserman, it puts MiniMax H3 / FastH3 video with synchronized sound, Qwen-Image-2.1 pictures, picture editing, Fish Audio Drama 3 voices, music, sound effects, an editor and Sam Wasserman’s Slate planner in one browser app with a queue. This guide launches the Studio (most easily from YouRunAI’s one-click RunPod template), opens it, and renders your first 5-second clip with sound from a text prompt. Locally it needs serious hardware: 96 GB of memory on a Mac, or a DGX Spark-class NVIDIA machine. Local speed and memory figures here are the publisher’s own measurements.

Choose software Run its commands Check the result

BEFORE YOU BEGIN

The software

Rent a GPU on RunPod (recommended) or install on a supported high-memory machine. Every route runs the same Studio and ends with the same first clip.

These requirements describe the GPU you rent. Use Windows, Mac or Linux as the browser client; your saved computer’s VRAM does not limit this setup.

CURRENT SETUP

Imagination Studio 4.3.1 · official

Needs 96 GB of memory (Mac Studio, DGX Spark or RTX PRO 6000), or a rented RTX PRO 6000 on RunPod.

View software source

THE WORKFLOW

Set up Imagination Studio 4.3.1 on your rented GPU

The steps below run on the rented Linux GPU; connect from your own browser. The instructions below update to match your selection.

CHOOSE YOUR SETUPRented GPU

Install guideRecommended

Rent a 96 GB RTX PRO 6000 on RunPod

The simplest way to run the full Studio without owning a supported machine. This is cloud GPU rental, not your own computer. YouRunAI generated images and FastH3 video with sound on a RunPod RTX PRO 6000, and a fresh Pod from the template finished its automatic setup and served the Studio.

GPU: 1 × RTX PRO 6000 Blackwell · 96 GB VRAM (rented by the hour)
Container disk: 500 GB or more
Your computer: Any computer with a web browser

RECOMMENDED · ONE-CLICK RUNPOD TEMPLATE

Launch Imagination Studio on RunPod

Installs Imagination Studio with FastH3 video with sound, images, music and voice. No install commands: deploy, wait for the first launch to finish, then open port 7080.

Affiliate link: YouRunAI may earn a commission. The template is YouRunAI’s deployment of the open-source Studio, not an official release by its creators.

Launch Imagination Studio on RunPod

Launch the template

Click Launch Imagination Studio on RunPod above and sign in; the template installs and starts the Studio for you. Select an RTX PRO 6000 Blackwell (96 GB VRAM) GPU, keep the container disk at 500 GB or more, and check the hourly price before you deploy. The models carry their own licences, including non-commercial and territory limits; check them before using clips commercially.

Setting	Value
GPU	1 × RTX PRO 6000 Blackwell · 96 GB VRAM
Container disk	500 GB or more
HTTP port	7080 (set by the template)

Launch Imagination Studio on RunPod Model licences and credits

Deploy and wait for the automatic setup

Deploy the Pod. The template installs Imagination Studio, FastH3 video with sound, image generation, music and voice by itself; there are no install commands to run. The first launch downloads and builds about 176 GB of models and engines, so it takes a while: follow progress in the Pod’s logs on RunPod, and wait until the Studio is up before opening it.

Open the Studio on port 7080

On the Pod, open Connect → HTTP Service on port 7080. RunPod’s address looks like https://YOUR_POD_ID-7080.proxy.runpod.net. If the Studio asks to sign this device in, open the Pod’s web terminal from Connect, run the command below and paste the key it prints. Treat the key as a password, and keep the sign-in on.

cat /root/ImaginationStudio/access-key

Make your first video with sound

Click Create in the sidebar. Set the five options below, paste the prompt as the scene, and queue the render. Progress shows in the Queue on the right. Once the models are loaded a clip like this takes about 1 to 1.5 minutes; the first one takes longer while they load. When the clip finishes, play it in the Library and download the MP4 from there.

Setting	Value
The prompt goes to	Video with sound
The clip starts from	Text only
Passes	8
Resolution	480p (864x480)
Length	5 s (124 frames)

Using the studio: Create

Plain text

A street musician plays a violin under a train bridge at dusk. Warm streetlights flicker on, a train rumbles overhead, and the camera slowly pushes in on her face as she finishes the phrase. Sound: violin, distant traffic, the train passing.

Download your files and stop the Pod

The Studio keeps everything under /root/ImaginationStudio on the Pod’s container disk, which does not survive a replaced or deleted Pod. Download the clips you want to keep from the Library (the simplest backup), then stop the Pod so the GPU stops billing. A separately configured network volume is the only RunPod storage that outlives a deleted Pod.

RunPod: storage types

Advanced: manual installation on a RunPod Pod

Use this only if you cannot use the template, or want to choose the install options yourself. It is the setup YouRunAI installed by hand on a RunPod RTX PRO 6000 Blackwell Server Edition: everything under /root on a 500 GB container disk (the tested /workspace mount gave Git and Python cache permission errors), port 7080, and a small patch so the Studio accepts jobs from the RunPod address. Run the commands in the Pod’s web terminal. When the Studio is running, make your first clip as in the main steps.

Deploy an RTX PRO 6000 Pod from a CUDA template

Sign in to RunPod and choose Pods → Deploy. Pick one RTX PRO 6000 Blackwell (96 GB VRAM) and a template with the CUDA toolkit; the tested Pod reported NVIDIA driver 595.91.07 and CUDA 13.2 (the Studio needs driver 580+ and nvcc 12.9 or newer). Edit the template: set Container Disk to 500 GB or more (the tested install downloaded about 176 GB plus build files and caches) and add 7080 under Expose HTTP Ports. Attach a network volume at /workspace only if you want somewhere durable to copy finished clips. Deploy, then open Connect → Web Terminal.

Setting	Value
GPU	1 × RTX PRO 6000 Blackwell · 96 GB
Container disk	500 GB or more
Expose HTTP ports	7080
Tested host	NVIDIA driver 595.91.07 · CUDA 13.2

Deploy on RunPod RunPod: expose ports RunPod: storage types

Install the tools and open tmux

The Pod runs as root, so no sudo is needed. nvidia-smi should show the RTX PRO 6000 and nvcc should print 12.9 or newer. Work inside tmux so the long install and the Studio keep running if the browser terminal disconnects: reconnect with tmux attach -t studio, and detach with Ctrl+B then D.

RTX Pro 6000 requirements

nvidia-smi
nvcc --version
apt-get update
apt-get install -y git curl ffmpeg tmux
tmux new -s studio

Read the model licences before the download

The install prints each licence link before it downloads. The defaults: Qwen-Image-2.1 is under the Qwen Research License (non-commercial); MiniMax H3 and FastH3 are under the MiniMax H3 Community License, which has territory limits. Optional parts add FLUX3-K (FLUX Kommunity License, non-commercial), Fish Audio Drama 3 and MiniMax-Music3 under their own community licences. The Studio’s code itself is Apache-2.0. Check these before using clips commercially.

Licences and credits

Set the install paths and clone the Studio

This saves the paths in /root/studio-env.sh so every later step uses the same install, models, caches and temporary files under /root, then clones the Studio. Run source /root/studio-env.sh again in any new terminal before using the Studio.

Imagination Studio repository

cat > /root/studio-env.sh <<'EOF'
export PREFIX=/root/tensorfold-studio
export SPARK_MODELS=/root/studio-models
export STUDIO_HOME=/root/ImaginationStudio
export HF_HOME=/root/.cache/huggingface
export DRAMA3_CUDA_MODEL=/root/studio-models/drama-3
export UV_CACHE_DIR=/root/.cache/uv
export XDG_CACHE_HOME=/root/.cache
export TMPDIR=/root/tmp
export PATH=/usr/local/cuda/bin:$PATH
EOF
mkdir -p /root/tmp /root/.cache/uv
source /root/studio-env.sh
cd /root
git clone https://github.com/drowzeys/Imagination-Studio-by-keys-Wasserman.git

Install images, video with sound, music and voice

The RTX Pro 6000 installer builds the CUDA engines for the card, downloads the models and finishes with a test image and a test clip with sound. --music adds MiniMax Music 3 and --voice adds Fish Audio Drama 3; leave them off for the core only. Read and accept each licence prompt. The tested run ended with DONE after downloading about 176 GB. If it stops, run the same command again; downloads resume.

RTX Pro 6000 installer Install flags

source /root/studio-env.sh
cd /root/Imagination-Studio-by-keys-Wasserman
bash oneshot-setup-rtx-6000.sh --music --voice

Let the Studio accept jobs from the RunPod address

Through RunPod’s proxy the page loads, but queuing a job fails with “Do this from the Studio page.” because the browser’s Origin is the proxy address. This patch lets the Studio accept that one exact address, read from STUDIO_PUBLIC_ORIGIN; it does not switch the Origin check off. It keeps a backup as app/server.py.bak, stops if the upstream code has changed, and the last line should print the patched line. Rerun it after updating the Studio with git pull.

cd /root/Imagination-Studio-by-keys-Wasserman
python3 - <<'PY'
from pathlib import Path
p = Path("app/server.py")
s = p.read_text()
old = 'elsewhere = origin is not None and urlsplit(origin).netloc.lower() != host.lower()'
new = '''elsewhere = (origin is not None
                     and urlsplit(origin).netloc.lower() != host.lower()
                     and origin != os.environ.get("STUDIO_PUBLIC_ORIGIN", ""))'''
if old in s:
    p.with_suffix(".py.bak").write_text(s)
    p.write_text(s.replace(old, new, 1))
elif 'STUDIO_PUBLIC_ORIGIN' in s:
    print("Fix already applied")
else:
    raise SystemExit("Upstream Origin code has changed; inspect before editing")
PY
python3 -m py_compile app/server.py
grep -n STUDIO_PUBLIC_ORIGIN app/server.py

Start the Studio on port 7080

RunPod normally sets RUNPOD_POD_ID inside the Pod; if the command stops with “Set your Pod ID first”, run export RUNPOD_POD_ID=your-pod-id and try again. STUDIO_ALLOWED_HOSTS lets the Studio answer on the proxy address and STUDIO_PUBLIC_ORIGIN lets it accept jobs from there. Leave this running in tmux. The tested launch also added STUDIO_ACCESS_KEY=off for a short session; that leaves the Studio open to anyone with the URL, so keep the access key on unless you are testing briefly and will stop the Pod afterwards.

source /root/studio-env.sh
cd /root/Imagination-Studio-by-keys-Wasserman
export RUNPOD_POD_ID="${RUNPOD_POD_ID:?Set your Pod ID first}"
export STUDIO_PUBLIC_ORIGIN="https://${RUNPOD_POD_ID}-7080.proxy.runpod.net"
export STUDIO_ALLOWED_HOSTS="${RUNPOD_POD_ID}-7080.proxy.runpod.net"
HOST=0.0.0.0 PORT=7080 bash scripts/app.sh

Open the Studio and sign in

Open https://YOUR_POD_ID-7080.proxy.runpod.net, or HTTP Service on port 7080 in the Pod’s Connect panel. When the Studio asks to sign this device in, paste the access key: open a second tmux window with Ctrl+B then C and run the command below. Treat the key as a password.

source /root/studio-env.sh
cd /root/Imagination-Studio-by-keys-Wasserman
bash scripts/studio-key.sh --key

Back up your clips, then stop the Pod

Everything the Studio makes lives under /root/ImaginationStudio on the container disk, which is lost if the Pod is replaced. If a network volume is attached at /workspace, copy your work there; a volume disk created with the Pod can be deleted with it, so download anything important to your own computer too. Stop the Pod when you finish so the GPU stops billing.

RunPod: storage types

mkdir -p /workspace/ImaginationStudio-backup
cp -r /root/ImaginationStudio/. /workspace/ImaginationStudio-backup/
ls /workspace/ImaginationStudio-backup/renders

MORE WAYS TO USE THIS SETUP

Add more of the Studio

The core install covers pictures and video with sound. Optional parts are flags on the same installer; on a Mac it only installs what the Mac’s memory runs.

Install every part this machine runs

--all adds every big model the machine supports: reference-to-video, FLUX3-K, MOSS-Image, MiniMax Music 3, Fish Audio Drama 3 voices, sound effects, the Nemotron 3.5 Bolt brain and Motion Previs (about 400 GB; 530 GB on a 128 GB+ Mac). FLUX3-K and sound effects are gated: accept their licences on Hugging Face and run hf auth login first. Use single flags such as --voice or --music to add one part.

Install flags

bash oneshot-setup.sh --all --app
# check what is installed, download nothing:
bash oneshot-setup.sh --verify

Update to a new version

Run this from the clone when no clip is rendering. It updates the engines and the parts already installed, fetches no new weights and leaves ~/ImaginationStudio (clips, projects, LoRAs, settings) alone. Restart the Studio afterwards.

git pull
bash oneshot-setup.sh --upgrade

MORE WAYS TO USE THIS SETUP

Plan a film in Slate inside the Studio

Sam Wasserman’s Slate is built in under Production. Its First AD sets up scenes and shots from a plain description, the Bible and Studios keep the look and cast consistent, and every shot renders through the same queue.

Turn a description into shots

Open Production → Wasserman Slate AI Studio and describe the film. The ✦ assistant buttons need a brain: your own Claude Code or Codex sign-in, or a local model (the Nemotron 3.5 Bolt brain installs with --bolt). Shots compile into prompts for the Studio’s own video engine, with a preflight that flags words the model will draw instead of avoid.

MORE WAYS TO USE THIS SETUP

Let a Hermes agent drive the Studio (MCP)

The Studio ships a one-file MCP server (Python 3.9+, no dependencies) that lets an agent list the Library, queue stills and clips, and follow the queue through the Studio’s own API. The Studio must be running. Hermes is YouRunAI’s standard agent harness.

Add the Studio to Hermes

Install Hermes Agent, then add this to ~/.hermes/config.yaml with the absolute path to mcp/studio-mcp.py in your clone, and start a new Hermes session. Inside the Studio, Brains & agents can also connect it for you.

Install Hermes Agent Studio MCP server

Plain text

mcp_servers:
  studio:
    command: "python3"
    args: ["/Users/YOU/Imagination-Studio-by-keys-Wasserman/mcp/studio-mcp.py"]
    env:
      STUDIO_URL: "http://127.0.0.1:7870"

Give Hermes a render task

Ask for something you can check in the Library. The agent uses the server’s queue_clip, job_status and list_library tools.

Plain text

Use the studio tools to queue three 5-second 480p clips with sound of the same lighthouse at dawn, noon and night, wait for each job to finish, then list the three files in the Library and tell me which one has the strongest sound design.

MORE WAYS TO USE THIS SETUP

Use the Studio from your phone

The Studio listens only on its own machine. To reach it from a phone or another computer without opening it to your network, use Tailscale and the Studio’s access key.

Sign a phone in

Install Tailscale on the Studio machine and the phone with the same account, run the command below once on the Studio machine, then open Settings → Access key in the Studio and scan its QR code with the phone. The key is a password to your Studio: share it only with your own devices.

Tailscale

tailscale serve --bg 7870

SUCCESS CHECK · Imagination Studio 4.3.1

A 5-second clip with sound plays in the Studio

The Queue shows the job finished, and the Library shows a 5-second 864x480 clip of the violinist under the bridge with the violin, traffic and train audible when you play it. The file is in ~/ImaginationStudio/renders on your own machine (inside Ubuntu on the N1X laptop) or /root/ImaginationStudio/renders on a RunPod Pod; on RunPod, download it from the Library. A local install sends nothing to a cloud service. To stop: run the app’s stop command on a Mac ("$HOME/Applications/Wasserman Slate AI Studio.app/Contents/MacOS/TensorFoldStudio" stop), press Ctrl-C in the scripts/app.sh terminal on Linux, close the launcher window on the N1X, or stop the Pod on RunPod after downloading your clips (its container disk does not survive a replaced or deleted Pod). To restart locally, open the app or run bash scripts/app.sh again; everything you made stays in ~/ImaginationStudio.

YouRunAI checked the local routes against the Studio’s repository (version 4.3.1, November 4, 2026); their hardware limits, timings and memory figures are the publisher’s measurements. YouRunAI ran the RunPod route itself. The app is unsigned and built locally from your clone; keep the clone where it is or rebuild the app with bash scripts/make-app.sh.

Back to setup steps

When it doesn’t go to plan

The installer stops and says the Mac has too little memory

Video, voice and music need 96 GB of unified memory. On a 48–95 GB Mac run bash oneshot-setup.sh --image-only --app for pictures only; below 48 GB the Studio is not supported.

“run oneshot-setup.sh first” or “no FastH3 checkpoint”

The install did not finish. Run the same install command again; downloads resume. bash oneshot-setup.sh --no-render finishes it without the test clip.

The app says it did not start

Read ~/Library/Logs/TensorFoldStudio.log. Often another program holds port 7870: rebuild the app on another port with PORT=7871 bash scripts/make-app.sh.

Renders are very slow on a Mac

Check that the chip is M5-family and nothing else is using the GPU. On M1 to M4, FastH3 runs on the slower MLX engine: use 480p, 5 s and 4 passes.

The job log says “MLX engine” for FastH3 on an M5

The native engine is missing or the clip uses a LoRA. Run bash oneshot-setup.sh --verify and look for the tf-h3-dit line; rerun the install if it is missing.

“This clip will be refused: memory” on a DGX Spark or GPU

Choose the size or length the message names, close other GPU programs, or free the card with bash spark/comfy.sh stop before a large job.

The installer stops because ffmpeg is missing (Linux)

Install it with sudo apt install -y ffmpeg and run the install again.

The engines do not build on an RTX 5090

The CUDA 12.8 nvcc fails. Install CUDA 12.9 or newer and set NVCC to its nvcc before running oneshot-setup-rtx-6000.sh.

A gated model is left out of the install

FLUX3-K and Stable Audio Open need their licences accepted on Hugging Face and hf auth login (or the N1X installer’s token prompt). The end of the install prints the command that adds them afterwards.

The N1X Studio stops when you close a terminal

WSL shuts Ubuntu down after its last Windows window closes. Start the Studio from the taskbar launcher and leave its minimized window open.

RunPod shows “Waiting for service to respond”

On a new template Pod the first launch is still downloading and building about 176 GB: watch the Pod’s logs and wait. After a manual install, check that port 7080 is under Expose HTTP Ports and that scripts/app.sh is still running in tmux (tmux attach -t studio). Then test the Studio locally with the proxy host name: curl -I -H "Host: ${RUNPOD_POD_ID}-7080.proxy.runpod.net" http://127.0.0.1:7080 should answer.

Pages load on RunPod but jobs fail with “Do this from the Studio page.” (manual install)

The Origin patch is missing or STUDIO_PUBLIC_ORIGIN is not set. Run grep -n STUDIO_PUBLIC_ORIGIN app/server.py in the clone; with no match, rerun the patch step, then restart the Studio with the launch command.

The RunPod page says to open the Studio by this computer’s name (manual install)

STUDIO_ALLOWED_HOSTS is not set to your Pod’s proxy address. Restart with the launch command from the advanced RunPod steps, which sets it from RUNPOD_POD_ID.

Permission errors from Git or Python caches under /workspace on RunPod (manual install)

Install under /root as the advanced RunPod steps do, with the cache and temporary folders from /root/studio-env.sh. Use /workspace only for backups of finished clips.

The Studio does not come back after restarting a RunPod Pod

Check the Pod’s logs first. After a manual install, background programs do not restart with the Pod: run source /root/studio-env.sh and the launch command again. If the Pod was replaced or deleted, its container disk was reset and your clips are gone unless you downloaded them or copied them to a network volume.

The RunPod Studio asks for an access key

Open the Pod’s web terminal from Connect, run cat /root/ImaginationStudio/access-key and paste the key. Do not turn the sign-in off on a Pod with a public address.

A job fails

Open its Log in the Queue; the last lines say why. Prompts with “no …” can draw the forbidden thing, because the model has no negative prompt: describe what should appear instead.

REFERENCE LIBRARY

Sources and files

Official software, installation instructions, and support resources behind this setup.

15 SOURCES

01Imagination Studio repository and READMEgithub.com 02Step-by-step tutorialgithub.com 03Install, and the one-click app (Mac)github.com 04Which Mac runs whatgithub.com 05Install flagsgithub.com 06DGX Spark (CUDA) buildgithub.com 07RTX Pro 6000 notesgithub.com 08RTX 5090 notes and measurementsgithub.com 09RTX Spark N1X laptop guidegithub.com 10Studio MCP servergithub.com 11Licences and creditsgithub.com 12Hermes Agent installationhermes-agent.nousresearch.com 13Imagination Studio RunPod template (YouRunAI)console.runpod.io 14RunPod: expose HTTP portsdocs.runpod.io 15YouRunAI: MiniMax H3 in ComfyUIyourunai.com

Keep exploring

Text

SETUP GUIDE

Qwen3.5 · 9B private chat setup

Source-backed package and platform-route plan for private local chat with Qwen3.5 · 9B. Preserves the official Ollama 9B package, a smaller 4B family fallback, and a clearly labeled community uncensored GGUF option.

Freshness 38 out of 100. Release age: 254 days old, Feb 27, 2026. Measures age, not quality.

Build a private coding assistant with Qwen3.8 27B

Run Qwen3.8 27B with Ollama, connect it to Cline in VS Code, and make one reviewable code change in a tiny local project.

Freshness 69 out of 100. Repository age: 95 days old, Aug 5, 2026. Measures age, not quality.

Arena Open Source Web coding: rank 12 of 62. Ranked model: Qwen3.8-27B. Original model ranking, not a benchmark of this package. Export dated Nov 6, 2026. Source checked Nov 7, 2026. Human preference ranking, not hardware suitability.+3

SOFTWAREOllama + Visual Studio Code + Cline + Python

Create images locally with FLUX.2 Klein

Generate images from text prompts in ComfyUI with the fast distilled or more adjustable base 4B model. Save your results and reuse the visual workflow.

Freshness 32 out of 100. Release age: 297 days old, Jan 15, 2026. Measures age, not quality.

SOFTWAREComfyUI

LOCALRENTED GPU

Verified source