Quality + UX release. No consensus or protocol changes — a drop-in upgrade for v0.10.3 miners.
Changes
- Possession-index progress counter — the model "staging / verifying" phase now logs live progress (chunks done, %, elapsed, ETA) every 10s, so a multi-minute index build on large models no longer looks like a hang.
- Cleaner OPoI answers — leading model reasoning / "thinking" blocks (Qwen3.5
</think>, GLM-4 and Gemma-4 channel-thought) are stripped from inference output before submission. - DRY anti-repetition sampler — added to the in-process llama.cpp engine's sampler chain; stops the small quantized models degenerating into whole-sentence loops that the flat repeat penalty missed.
Builds
| Line | CUDA | Driver floor | GPU archs |
|---|---|---|---|
modern (keryx-miner-supr-0.10.4*)
| 12.9 | 575+ | sm_75–120 (Turing → Blackwell) |
legacy (keryx-miner-supr-legacy-0.10.4*)
| 12.4 | 535+ | sm_70+ (Volta / older-driver rigs) |
Each line ships a HiveOS package (*.tar.gz) and a Linux package (*-linux-x86_64.tar.gz), and bundles both AVX and noavx llama CPU backends (auto-selected by CPU capability). AMD and Windows builds are provided as the *-amd-* / *-windows-* assets.
Checksums: SHA256SUMS-0.10.4.txt (NVIDIA Linux/HiveOS), SHA256SUMS-amd.txt (AMD).