v0.9.3 — --model-dir: shared/custom model directory
New option for rigs that keep the OPoI models on a shared or custom path:
--model-dir <DIR> Directory to look for AND download the OPoI models
(default: 'models' next to the miner binary)
- The miner looks up models in
<DIR>/<Model-Name>/model.ggufand saves downloads there too — point several rigs at one network share and download each model once instead of per-rig (the tiers are 6-28 GB). - The per-model possession-tree cache (
pom-tree.bin) lives in the same directory, so a shared store also skips the one-time index build on every rig after the first. - The path is validated at startup: created if missing, non-directories rejected,
~/expanded. A read-only share is allowed for pre-staged models (the miner warns that downloads of missing models would fail).
Also in this release:
- README: new "Model directory" section; tier table fixed for H5 (tier-0 = Qwen3-8B-abliterated, EXAONE retired).
- All five tiers re-validated live on the H5.2 chain (RTX 5090, krx.suprnova.cc): light / default / high / very-high all mine with 0 rejected shares. If a tier looks "broken" on your rig, the usual cause is the model still downloading over IPFS (PoW stays gated until the GGUF is complete) — pre-stage the model into your
--model-dirto skip the wait.
Consensus behavior is unchanged from v0.9.2 (H5.2 era, keryxd v1.4.0).
Asset repack (2026-07-26): the Linux/NVIDIA tarballs above were refreshed in place with a new --hiveos option: on HiveOS the bundled h-run.sh now passes it automatically, defaulting the model directory to /hive/miners/custom/models — outside the versioned miner folder — so models survive miner upgrades instead of re-downloading 6-28 GB each update. An explicit --model-dir overrides it; non-HiveOS boxes fall back to the old behavior with a warning. Re-download the tarball (SHA256SUMS updated) to pick it up; binaries still report 0.9.3.