Observed arrival · 2026-09-05
Xyntetik Runner Wants Local Inference You Can Prove
An Apache-licensed, single-binary C engine for serving GGUF models, training LoRA adapters, and recording locally run inference as replayable evidence.
Field notes
The homepage describes Runner as a plain-C, single-binary engine for GGUF models across CPU, CUDA, and Metal, with no Python or pip layer. Its evidence pages focus on failure modes rather than headline speed: one example covers seven token budgets across six engines, while another reports training a Qwen3-4B Q4_K_M adapter through the deployed 4-bit weights and reproducing the same SHA-256 twice. The page also describes signed receipts for replayable runs.
Observed signals
Read the marks
Editorial observations of this landing page, not a rating.
One card from the complete issue