| | Add Gyro-L (64 GB tier): TQP pair-trellis experts, 50.2 GiB GPU weights | agentionai | Oct 12, 2026 |
| | Gyro-M update: improved encode (KLD 0.291) | agentionai |
| | Card: point Strata instructions at rc2 | agentionai |
| | Gyro-S update: improved encode (KLD 0.401) | agentionai |
| | Docker Hub image, sponsor link; note on multi-request MTP fix | agentionai |
| | MTP draft: upstream tensor names (blk.48.nextn.hc_head_*) + MTP-layer compress ratio; weights unchanged. Loads in upstream llama.cpp b11386+ and agentionai/llama.cpp main 2026-10-05+ | agentionai |
| | Card: Windows note for --ngram-on-disk (fast reader on Windows from main 2026-10-05; lazy-mode fallback) | agentionai |
| | Card: corrections (sizes, loop results over 10 seeds, Vulkan or CUDA, 5090 session 2 numbers, Strata quick start, MMLU-Pro) | agentionai |
| | Card: 16-24 GB cards supported with --n-cpu-moe (new CPU kernels, 2026-10-02) | agentionai |
| | Card: NVIDIA RTX 5090 results (CUDA) and Vulkan-on-NVIDIA container notes | agentionai |
| | Add Gyro-M (48 GB tier): 2.125-bit experts, 8-bit shared layers | agentionai |
| | Gyro-S replaces the first Gyro file; card becomes the S/M/L family card | agentionai |
| | Update README.md | agentionai |
| | initial commit | agentionai |