Bittensor subnet Ingot beats Qwen's own FP8 quantization within 4 days
const_reborn · x · 2026-10-19
Ingot (SN97), live for only 4 days, already produces a better quantization of Qwen3.8-27B than Qwen's own FP8 release — 34% closer to the original model at the same size, FP8 only, with lower bits coming next.
Miners quantize the teacher to FP8, and every copy is assayed against the original token-by-token and layer-by-layer; the closest one is stamped king and earns rewards. A chart shows four kings in four days, fidelity climbing from 0% to 34%.
More from Infra
- AMD, NVIDIA, and Google All Ship Major vLLM Inference Optimizations in Three Days — PyTorch · 2026-10-20
- XFreeze Claims SpaceX Orbital Data Centers Are the Endgame for Scaling Intelligence — XFreeze · 2026-10-20
- Intel's Crescent Island GPU D270: A 160GB PCIe 5.0 Inference Card — lemire · 2026-10-20
- Mirai's uzu hits 123 tok/s decode on Apple M5 Max 128GB, fastest for Qwen-3.8 27B — teortaxesTex · 2026-10-20
- turbopuffer cuts disk read I/O by up to 80% with smarter vector rescoring — DanielLockyer · 2026-10-20
- vLLM Speculators v0.9.0: 2x faster speculative decoding training, 50% less storage — vllm_project · 2026-10-20