Run a JevK5 miner
The hosted training queue supports JevK5 4B today. Approved miners train candidate adapters; validators evaluate them; trusted Zils inference servers serve shared and accepted customer models.
Choose the right workflow
| Workflow | Model | Guide |
|---|---|---|
| Hosted queue miner and evaluator | JevK5 4B | JevK5 setup |
| Customer upload and approved-miner access | Model pinned per job | Shared queue |
| Capacity-gated assignment and accepted-model activation | JevK5 4B | Automatic workflow |
| Local bundled fleet and Bittensor testnet | Kev 0.8B | Legacy miner guide |
Prepare an approved GPU miner
- Install the pinned dependencies and create the verified model reference using JevK5 training setup. Use Linux or WSL 2, Python 3.13, and a BF16-capable CUDA GPU.
- Register the miner and create its protected hotkey/wallet configuration using shared queue setup. For the hosted service, set the miner configuration’s
coordinatortohttps://training.zils.ai. Miners receive training exports only; keep service credentials and evaluation data with the coordinator and validator. - Set the active model and start the matching miner and processor using service configuration. Shared hardware needs the documented compute lock and capacity checks.
- Configure the automatic workflow when you want approved assignment and verified activation of accepted customer models.
The hosted training model is jevk5-4b-v0.3. Its candidates contain adapter_config.json, adapter_model.safetensors, and model.json; accepted downloads add release.json. Existing Kev jobs retain their original identity and artifact format.
Training and inference are separate responsibilities
Miners do not serve customer prediction traffic. The decision API authenticates callers and routes model requests to trusted runtimes. The workflow activates a private customer model only after acceptance and verification of the runtime identity.
no_qualifying_model is a valid training completion. A rejected candidate does not replace the shared model or the customer's current version. Version selection defines first-version and upgrade comparisons.
Read the evidence in scope
The older testnet round and Kev queue pilot keep their original model names. They are historical evidence, not descriptions of today's hosted model. The research record separates model-quality experiments from workflow checks.