Windows
Windows 10/11 · x64
This device does not appear to be x64. The installer for this platform is x64 only.
Download .exe46.9 MiBFROM DOWNLOAD TO A TIMELINE
Your first shot, and the timeline it lands on.
The install takes a few minutes. The rest of the first session goes to connecting execution you already pay for, making one shot, and sending it somewhere an editor can open.
KUMIHO DESKTOP
v0.4.0Get Desktop for your operating system
Your detected platform is highlighted in fox orange. Every other option remains available.
Choose Windows, macOS, or Linux.
Detected
macOS
macOS 12+ · ARM64 / x64
Linux
Linux · x86_64
This device does not appear to be x64. The installer for this platform is x64 only.
WHAT YOU BRING
9Miho ships no model and no generation quota.
It boots without any generation credentials at all — install it, look around, and connect execution when you want to run something. There are three ways to do that, and they are not exclusive.
A CUDA GPU and a ComfyUI-backed stack
Nothing leaves the machine and there is no per-generation charge. It costs time instead: on an RTX 3080, MiniMax H3 at 864x480 for five seconds is about seven minutes.
An account you already have
9Miho stores the provider profile and the provider bills you directly. Seconds to minutes per shot, and the fast frontier models live here.
A GPU you start and stop
Point 9Miho at a ComfyUI endpoint you bring up when a run is queued and shut down when it is not. Local speed without leaving a card idle.
Model availability follows your provider account, region, entitlement, and configured endpoint. 9Miho does not silently substitute a different model when the one you asked for is unavailable.
THE FIRST SESSION
Five steps, and two of them are the installer.
- 01
Get Kumiho Desktop
Download the Desktop release for your platform. No source access, no GitHub sign-in, no toolchain.
- 02
Install 9Miho
Open the 9Miho card and choose Install. Desktop picks and verifies the platform bundle and keeps it current.
- 03
Connect execution
Add a provider profile, or point at a local or on-demand ComfyUI endpoint. This is the step that decides what a run costs you.
- 04
Make one shot
In Canvas: pin a reference, write a Direction, run it. The output registers as a revision of an asset the moment it exists, so the shot you approve stays the shot you approved.
- 05
Send it out
In Cutting Room: colour-managed review, then an OTIO timeline with a 10-bit mezzanine and a per-clip CDL. That is the part the edit suite can open.
Available forWindows · macOS · LinuxChoose an installer by platform ↑
The same front door after install.
No source access required
A released Kumiho Desktop installs the bundled runtime without access to the private source repository or a GitHub login.
No development toolchain
A normal install does not require Python, Node, or package-manager setup.
No version chasing
Desktop checks 9Miho’s public release feed and only offers a version newer than the one installed.
No project replacement
An update replaces and restarts the runtime while preserving existing Projects, Assets, Data, and Logs by design.
LOCAL + HOSTED · ONE PRODUCTION
Local GPU or paid API. All three rooms on one production engine.
Use supported ComfyUI-backed local or self-hosted model stacks, on-demand endpoints, and paid provider APIs through one typed production engine. Instead of rebuilding a low-level ComfyUI graph for every shot, choose the exact capability and Variant.
SUPPORTED MODEL FAMILIESChoose the right model for each shot.
The support catalog in the current public release. Showcase credits name the model actually used for each reel.
MiniMax
MODEL FAMILY- Video
- Music
- MiniMax H3 · APIProvider API
- MiniMax H3 · ComfyUIComfyUI service · local / on-demand
- MiniMax H3 TurboComfyUI service · local / on-demand
- MiniMax H3 Reference HQ · Two-StageComfyUI service · local / on-demand · conditionalRequires NVIDIA VFX; disabled in the default on-demand profile
- MiniMax H3 Sage + Sol · 15+5 Latent UpscaleComfyUI service · local / on-demand · conditionalRequires a separately provisioned, canary-verified Sage + Sol runtime; disabled in the default on-demand profile
- MiniMax Music 3.0 / 3.0 Free / 2.6 / 2.6 FreeProvider API
- MiniMax-Music3 · ComfyUIComfyUI service · local / on-demand
ByteDance Seed
MODEL FAMILY- Video
- Image
- Seedance 2.5Provider API
- Seedance 2.0 / Fast / MiniProvider API
- Seedream 5.0 Lite / ProProvider API
- Video
- Image
- Analysis
- Veo 3.1 / Fast / LiteProvider API
- Nano Banana 1 / 2 / ProProvider API
- Gemini Omni FlashProvider API
- Gemini 2.5 FlashProvider API
Kling AI
MODEL FAMILY- Video
- Kling 3.0Provider API
- Kling 3.0 OmniProvider API
Black Forest Labs
MODEL FAMILY- Video
- Image
- FLUX.2 [flex] / [max] / [pro]Provider API
- FLUX 3 VideoProvider API
OpenAI
MODEL FAMILY- Image
- GPT Image 2Provider API
- GPT Image 1Provider API
View 6 more model families
- SeedVR2SeedVR2 7B FP16Video / Image / Finishing
- KreaKrea 2 Turbo FP8, Krea 2 Identity Edit V1.2Image
- Lightricks · LTXLTX-2.5, LTX-2.3, LTX-2.3 · AI HDR remaster · conditionalVideo / Finishing
- Alibaba Cloud · WanWan 3.0Video
- ElevenLabsElevenLabs Speech, ElevenLabs Music, ElevenLabs Sound EffectsVoice · TTS / Music / SFX
- Topaz LabsImage Standard V2, Video ProteusVideo / Image / Finishing
Revisioned media library
Reuse a pinned image, video, or audio revision without a later “latest” result silently changing the authored graph.
Traceable results
Inspect the producing run, exact inputs, actual rendered prompt, selected tier, validated settings, and redacted production record.
Agent-proposed. Checked before generation.
The agent proposes character-state changes and scene direction; 9Miho checks story time, state handoff, and references before generation. The creator decides what enters canon and whether a paid run proceeds.
SHARED FOUNDATIONKumiho graph
Each product has a clear production role. Assets keep their history, agents carry context forward, production workflows coordinate the work, and 9Miho turns it into images and video. These connections are being released progressively.
Kumiho
Creative asset systemStable identity, revisions, Artifacts, relationships, and provenance for production assets.
Kumiho Memory
Agent memory layerWorking memory, consolidation, recall, and typed knowledge that can carry evolving story context forward. Story specialization is part of the ecosystem direction.
Kumiho Desktop
The front doorThe native place to choose a Kumiho connection and install, launch, inspect, and update 9Miho.
Revka
Production orchestrationAn independent runtime for agents, teams, workflows, tools, approvals, and operating evidence. 9Miho integration is progressive.
9Miho - Canvas
Visual productionThe Canvas engine turns approved story context into media; Cutting Room edits and colour-grades the selected takes into an episode, then renders the final cut or hands it to a master suite.
9Miho - Storyteller
Narrative canonApproved character state, the event that changed it, and the state the next Moment reads. It exchanges that story context with the rest of the graph.
ComfyUI can remain the execution backend for applicable local model families. 9Miho bundles no model weights. Actual availability depends on the configured endpoint, account, region, entitlement, and local hardware. Product names and marks belong to their respective owners.
Let the next scene remember the last one.
The first shot can be luck. Production starts at the second one, and everything that makes the second one possible has to be in place before it exists.