Stable Horde

Where Stable Horde sits

Stable Horde — the project now styles itself AI Horde, run by the Haidra organization — is not a startup product and should not be judged like one. It is a volunteer computing network in the Folding@home and SETI@home tradition: people point their idle GPUs at a shared queue, strangers' prompts get generated on that hardware, and nobody charges anybody money. The stated mission is explicit about the politics of it — universal access with no paywall, priority that can only be earned by contributing compute rather than purchased, and a registered non-profit structure. Whether you find that charming or impractical determines most of your experience, because every trade-off in the product flows from refusing to sell what everyone else in this category sells.

What I look for in a generation service is boring and commercial: does it produce usable output, how long do I wait, and what recourse do I have when it breaks? On a paid service those three questions have crisp answers. On the Horde they have honest ones.

What it actually does

The system is a queue in front of a distributed pool of volunteer workers. You submit a request through the REST API or through one of the community clients — ArtBot is the polished web front end, plus Discord bots and assorted apps — and the server validates your parameters, splits the request into jobs, and hands each job to the first suitable worker by priority order. Three worker types exist: Dreamers generate images with Stable Diffusion, SDXL, Flux, and similar open models; Scribes run text generation through LLM backends such as KoboldCpp and Aphrodite; Alchemists do post-processing — upscaling, face fixing, captioning, interrogation — and several of those tasks run on a plain CPU, which means you can contribute without a graphics card. Workers poll outbound for jobs and never listen on ports, requests are anonymized before dispatch, and a trust-and-suspicion system flags workers who produce bad output and eventually excludes them.

The economy is kudos, and the design is the interesting part. Every job is priced in kudos by a trained neural network that estimates generation time against a baseline of 10 kudos for a 50-step 512x512 image — a job expected to take twice that is 20. Your kudos balance sets your position in the queue; anonymous users (the API key is ten zeros) get lowest priority, registered users start with 25 kudos, and workers earn kudos per completed job plus a small uptime grant every 10 minutes just for being available. A small horde tax is burned on every request (1 kudos per request, 1 per job, plus 20% extra burn if you demand fast workers only and 10% for a worker blacklist) to fund the infrastructure layer. Kudos never expire. They can be gifted or transferred between users freely, but selling them is a terms-of-service violation that can zero both accounts — the project treats commodification as an existential threat, and the mission page is blunt that priority can only be earned with compute. Most earned kudos go unspent, which in practice means the volunteer pool subsidizes anonymous newcomers as a matter of course.

There is no pricing page because there are no prices. The service is free to use, anonymously if you like. Financial donations to the project exist and come with a kudos thank-you, offset by kudos minted on core-developer workers so that donating does not buy queue position.

My testing notes

Two observations from working with the queue, plus one design note. First: the honest headline is that it works and it costs nothing, and the waiting is the price. Under normal load a request comes back in a timeframe that feels like a slow paid API; under heavy load, requests can time out and the project's own FAQ advice is to simply retry, possibly with a different model or easier parameters. That is not a defect they are hiding — the FAQ says plainly that most fulfilled requests come from anonymous users — but it means the product has no latency contract at all. If you need images on a deadline, this is not your primary service. If you are a student, a hobbyist, or someone exploring open models without a GPU, it is remarkable what you get for zero money.

Second, the quality floor is set by the worker pool and the trust system, not by a central provider. Outputs from the same prompt vary more than on a hosted service, because a volunteer might be running a model checkpoint and settings that suit their own taste. The trust-and-suspicion mechanism and the model selector are your tools here: pick a popular model with many workers behind it and results stabilize. The parameter validation is strict — requests outside the published constraints get rejected with the offending field named, and resolution thresholds can require a kudos balance or an allow-downgrade adjustment — which keeps the worst failure modes off the queue. The dry-run parameter is a small kindness worth using: it returns the kudos cost of an image request without generating anything, so you can price a batch before committing your queue position.

The design note: the kudos system is the most thoughtful anti-commodification mechanism I have seen in this space, and it genuinely works as intended. You cannot buy priority with money, full stop, and the option to bank kudos by running a worker when you do not need it — then spend it when you do — turns an idle GPU into a form of insurance. Alchemy jobs on 2GB of VRAM or a CPU give the hardware-poor a path to earning priority too. The trade is that the system deliberately refuses to grow a paying customer class, so there is no funded support tier, no SLA, and no recourse beyond retry and community help.

What it does not do

There is no service-level agreement, no guaranteed latency, and no guaranteed output quality — the trust system polices the worst offenders but variance is structural. There is no human support desk; help comes through the Discord community and documentation. Commercial use of outputs depends on the license of whichever open model generated them, and that varies (some Flux and SDXL checkpoints carry their own terms), so the Horde's free price does not settle your rights to what you make. Text generation is limited to whatever models Scribes choose to serve, which is a narrower and more variable catalogue than a hosted LLM subscription. NSFW filtering is available per request and illegal-content checks always run, but you are generating on volunteer machines, so the moderation posture is per-worker. And because kudos cannot be sold or bought, there is no way to throw money at a slow queue when you are in a hurry.

Verdict

Judged as what it is — a mutual-aid compute cooperative rather than a vendor — the Horde is one of the more admirable projects in this category, and the fact that anonymous users routinely get free generations funded by volunteers' unspent kudos is quietly extraordinary. I rate it 3.6, scoring free access and a genuinely principled kudos economy strongly on value, a functional if variable generation pipeline on core functionality, simple client-side usability offset by queue unpredictability, a mature open-source and community ecosystem (public API, GitHub, Discord), and reliability signals limited by the honest absence of any contract between you and the service. Use it to learn, to experiment with open models, or to generate without a GPU — and keep a paid fallback for anything with a deadline.

Sources

• https://www.stablehorde.net/
• https://stablehorde.net/mission/
• https://stablehorde.net/contribute/workers/
• https://horde-sdk.readthedocs.io/en/latest/haidra-assets/docs/kudos/
• https://haidra-org.github.io/AI-Horde/haidra-assets/docs/workers/
• https://github.com/Haidra-Org/AI-Horde/blob/main/FAQ.md
• https://aihorde.net/faq/
• https://aihorde.net/
• https://github.com/Haidra-Org/AI-Horde
• https://stablehorde.net/details/workers/