| | Update README.md | agentionai | Sep 14, 2026 |
| | Update README.md | agentionai |
| | Update README.md | agentionai |
| | Credit charlie12345 as the origin of the ROCmFP4 format | agentionai |
| | card: point at the renamed model file | agentionai |
| | rename: drop 'imatrix' from the model filename so -hf can resolve it | agentionai |
| | card: adaptive MTP + q8_0 KV run instructions, refresh throughput figures | agentionai |
| | Move throughput chart to the top of the card | agentionai |
| | Note build date and that generation figures exclude MTP | agentionai |
| | Add long-context throughput section with depth-sweep chart | agentionai |
| | Add throughput-vs-depth chart | agentionai |
| | Fix mmproj path; add Agention link | agentionai |
| | Move mmproj into mmproj/ so HF indexes the text model, not the vision tower | agentionai |
| | Rewrite card: benefits list, prefill/PPL comparison, vision + MTP usage, sampling params; drop v1 refs | agentionai |
| | Remove superseded v1 build (5-shard root + joined/), 170 GiB; v2 supersedes it in both layouts | agentionai |
| | Add vision tower (mmproj, f16) - verified with llama-server | agentionai |
| | Drop specific tensor size, keep the buffer-limit note general | agentionai |
| | Fix size: both layouts are 87.06 GiB (GiB/GB unit mix-up on an identical byte count) | agentionai |
| | Add v2 per-head PLE layout (VRAM-resident) | agentionai |
| | Add v2: BF16 source + Q6_K/Q8_0 backbone, 4.1062 PPL | agentionai |