📦 Which file do I download?
| Setup | File |
|---|---|
| NVIDIA — HiveOS | keryx-miner-supr-0.10.14.tar.gz
|
| NVIDIA — mmpOS | keryx-miner-supr-mmpos_0.10.14.tar.gz
|
| NVIDIA — Linux | keryx-miner-supr-0.10.14-linux-x86_64.tar.gz
|
| NVIDIA — Windows | keryx-miner-supr-windows-nvidia-pom.zip
|
| AMD | keryx-miner-supr-amd-* / keryx-miner-supr-windows-amd.zip
|
-legacy = older drivers (~535)/older cards; plain = modern (575+). Verify with SHA256SUMS-0.10.14.txt.
v0.10.14 — multi-GPU crash fix + native sm_120 Windows
Fixes the multi-GPU crash. On a multi-GPU rig the miner could connect, download the model, then crash silently at the walk build (single-GPU rigs and the official miner were unaffected). Cause: the PoM walk didn't bind each device's CUDA context to its worker thread, so on a multi-GPU host a raw kernel launch/memcpy could run against another card's context → CUDA_ERROR_ILLEGAL_ADDRESS → silent native crash. Now every walk entry point binds the device context first (matching the reference miner). Applies to Windows and Linux multi-GPU rigs.
Native sm_120 on Windows. The Windows keryx-llama.dll is now built with native Blackwell (5070Ti/5080/5090) SASS (120-real, CUDA 12.8.1) instead of PTX-JIT only — hardens Blackwell inference on Windows.
Also in the recent 0.10.x line: H8 request-identity for solo miners (v0.10.13, mandatory before DAA 79,251,000), Windows silent-stall/no-download fix (v0.10.12), escrow claim-batch fix, periodic capability re-declare, loud model staging, Gemma-fits-16GB.
Assets: HiveOS/mmpOS/Linux (modern + legacy) + Windows (NVIDIA + AMD) + SHA256SUMS. AMD attached separately.