There's a peculiar ambition brewing on Subnet 99 of the Bittensor network. Where OpenAI guards Sora behind waitlists and API keys, and Google dangles Veo as the next frontier in generative media, a scrappier project is attempting something rather different: open-source AI video generation powered not by Big Tech's server farms but by a scattered network of miners running GPUs in basements, data centers, and wherever else the compute happens to live.
The project is called Leoma. Built by Rendix Network, it launched a mining competition in April that turns AI video generation into a kind of distributed race—participants stake compute power, submit clips, and get ranked by how well their outputs match the prompts validators throw at them. The code is open, available under the MIT license on GitHub. The models are public, uploaded to Hugging Face. And the underlying bet is that decentralized incentives can produce results competitive with companies that burn through billions in venture funding.
Whether that's realistic or wildly optimistic depends largely on who you ask.
Mining for Video, Not Just Coins
Leoma operates as a Text-Image-to-Video platform, which in plain terms means miners receive a text prompt and an initial frame, then return a short video clip. The system evaluates those clips—currently using OpenAI's GPT-4o, an irony not lost on decentralization purists—scoring them on fidelity, motion consistency, temporal stability, visual quality, and how faithfully they track the prompt. Those scores feed into on-chain weight adjustments each epoch. Perform well, earn more emissions. Lag behind, and your share diminishes.
Right now, the platform handles only TI2V. Pure text-to-video and image-to-video modes exist on the roadmap, mentioned in documentation updated mid-April, though neither has gone live yet. The models miners deploy are variations of Wan2.2, a roughly 14-billion-parameter architecture that gets fine-tuned and uploaded to Hugging Face under a strict naming convention: repositories must start with "leoma" and end with the miner's unique hotkey.
It's a system that tries to solve a familiar tension in decentralized AI—how do you incentivize quality when the network is permissionless and the outputs subjective? Leoma's answer involves automated scoring, public leaderboards, and on-chain weight adjustments that reward miners who produce better outputs and penalize those who fall behind.
The Hardware Hurdle
Getting started as a Leoma miner isn't a casual weekend project. For 480p video output, you'll need at minimum an RTX 4090 with 24GB of VRAM. Want to handle 720p? That requires an A100 with 40GB. Add to that 200GB of SSD storage, 32GB of RAM, and an 8-core CPU. Validators—the nodes that issue prompts and score results—face even steeper demands: A100-class GPUs, 500GB of storage, 64GB of RAM, 16 cores, and network bandwidth north of 100 Mbps.
The documentation, published under an MIT license, walks through Docker setups, environment variables, and object storage backends. Leoma supports S3-compatible storage via Cloudflare R2 or Hippius, another Bittensor subnet (Subnet 75) that offers decentralized alternatives to AWS. Validators pull tasks from storage, run evaluations, and POST results to the Leoma API. Each epoch, they set weights on-chain using the subnet's metagraph.
It's a carefully architected stack designed to avoid centralized bottlenecks, though the reliance on GPT-4o for evaluation raises an obvious question: if you're leaning on OpenAI to judge video quality, how decentralized is the system really?
A Rebrand and a Moving Target
Subnet 99 wasn't always called Leoma. Until recently, it went by Neza. Some third-party tracking sites still list it under the old name, referencing agent-based workflows or earlier model versions like WAN2.1. The rebrand to Leoma appears to have coincided with the April launch, though documentation around the transition is sparse. TaoStats and SubnetRadar now track it as Leoma; a snapshot from SubnetRadar showed the subnet's alpha token priced around $1.56 with a market cap hovering near $777,000—numbers that shift constantly and should be taken as historical rather than current.
Social media posts from the Leoma account have floated the idea of training models at 25 billion parameters or beyond. That claim drew skeptical responses on Reddit, where users questioned whether a decentralized subnet could realistically compete with companies that have vastly larger budgets, exclusive access to cutting-edge hardware, and years of accumulated expertise. The counterargument—embedded in Leoma's very design—is that distributed networks can aggregate resources and iterate faster than monolithic entities. Miners fine-tune models, validators score outputs multiple times per epoch, and the entire system updates on-chain.
Whether that velocity translates into quality that can genuinely rival Sora or Veo is an open question. Perhaps more than the founders anticipated.
The Bigger Bet
Leoma's pitch is disarmingly straightforward: instead of waiting months for access to Sora or paying for Runway's API, developers and creators can tap into a decentralized network where the models are open, the code is public, and the incentives theoretically align around producing better outputs. It's a vision that resonates in certain corners of the tech world—especially among those who view the consolidation of AI capabilities in a handful of companies as a structural risk.
But vision and execution are different beasts. The mining competition is live, the code repositories are public (available at docs.leoma.ai and github.com/RendixNetwork/leoma), and early outputs are flowing through the network. What's less clear is whether those outputs can close the gap between "interesting experiment" and "viable alternative to commercial platforms."
Leoma has published a whitepaper and litepaper, though access to the PDFs wasn't straightforward during research for this piece. The roadmap includes expanding beyond TI2V to support pure text-to-video and image-to-video tasks. For now, the focus is onboarding miners, refining the evaluation pipeline, and proving that the subnet can produce clips worth the compute being thrown at them.
The subnet operates on Bittensor's Finney mainnet as Subnet 99, capped at 256 UIDs and 64 validators. The competition is live. The barriers to entry are high but not insurmountable for developers with access to serious GPUs and an appetite for experimentation.
For an ecosystem built on the premise of decentralizing machine intelligence, Leoma represents a test case that goes beyond technical feasibility. It's a bet that distributed incentives can drive the kind of model improvement that has historically required billions in venture capital and exclusive access to TPU clusters. The code is open, the first outputs are already circulating, and the network is running.
Whether it works—really works, in a way that matters outside of crypto-native circles—is a question that won't be answered in a single epoch. But the experiment is underway, and for now, that might be enough.
