Greedy soup.
The best z3 isn’t one model. It’s a bowl.
Every tier is a soup. Up to five bodies, from that tier’s open house models and whatever miners bring, averaged into one model and tasted again every day. A body keeps its seat only while the soup is better with it.
| Soup seatsFive fixed seats. Each member whose contribution is at least 0.8 points earns one seat, times its rate. A seat is worth the same however many others are seated. | 5 × 16% |
|---|---|
| Solo crownThe round’s best single body that is not a house model, times its rate, only if it beats every house model and the current soup by at least 1.2 points. A soup usually beats its members, so expect the crown to burn in most rounds. | 20% |
| Anything unearnedEmpty seats and an unclaimed crown pay no one. | Burned |
| TrainerYou trained the body, well away from every listed public model. Fine-tunes of a house model, the soup or a release count; at launch their seats pay only if they beat what they started from by at least 1.2 points. | 1× |
|---|---|
| ScoutAn open model you found and committed first. One on the tier’s known-public list, re-uploaded as is or lightly changed, pays 0.25×, as do fine-tunes of a listed model at launch. A find that is not on the list and well away from every listed model pays the full trainer rate. | 0.25× to 1× |
| House mixA body that is almost entirely a blend of house models and listed public models, or of these with the soup and the release, with less than 5% of its own change. | 0.25× |
| House modelThe open models each tier’s soup starts from. They cook; they are not paid. | 0 |
The stock is already on.
Base and house models
Five open models already fine-tuned from that base go into the pool every round. They are the free starting point and are never paid. Your body joins only if it makes the soup better.
- d3-miniby vllm-sr · Also z3-mini v1, the founding baseline
- pplx-decider-flash-v1.2by perplexity-ai
- torchcast-decision-4bby torchcast-ai
- Matilda-Decision-v1-Miniby Maincode
- KnowLine-4B-Gen5by PelaAI
The house models make sure there is always a soup to beat; miners earn by making it better.
- Any fine-tune of your tier’s Qwen base (
Qwen/Qwen3.5-0.8Bfor lite,Qwen/Qwen3.5-4Bfor mini,Qwen/Qwen3.5-9Bfor flash,Qwen/Qwen3.8-27Bfor z3) with that tier’s body layout, trained from plain Qwen, a z3 release, the current soup or another open model. - You bring the body only. The tier’s answer layer, prompt and runtime stay fixed.
- Move at most about 15% away from Qwen.
- Apache-2.0, MIT or BSD licence.
- Your repo names your hotkey in
zils-owner.json. The first valid commitment of a body wins it. - Scouts welcome. Any open fine-tune that passes these rules can be committed by whoever finds it first, listed or not. The known-public list only marks the models validators already know about; it is not the limit of what you can bring.
- No reheats: near-exact re-uploads, bodies built largely from other miners’ bodies, and models that answer like a pool body are turned away.
- Up to 20 bodies compete for the five seats each round. At most 6 new bodies are admitted per round, in commitment order. Flash and z3 take fewer.
- Upload to a private repo, commit, then make it public before the day’s commit deadline, so nobody can commit your body first.
A good soup becomes the next release.
Every 14 to 28 days the soup can become the next release. It must beat the current release by at least one point on two sealed release exams that rounds never use, hold that lead 12 hours later, and keep customer adapters at least as good. A reviewed promotion then makes it that tier’s next version. Its members are paid through their seats; the release itself has no single winner. Every tier is a soup with its own house models, and each is released on its own. Lite and mini cook from the launch; flash and z3 open later.
Pull up a stool.
curl -fsSL https://zils.ai/install.sh | shUpload your body to Hugging Face, then check and commit it with zils-miner commit-body. Without --yes it is a dry run that prints the commitment and writes nothing. A hotkey has one commitment slot, shared with base candidates.
Training needs Linux x86_64 with an NVIDIA GPU. Read the miner guide for body rules, committing and pay. These numbers are launch policy and may change.