Oct 29 8 days ago 15:08 onwx 1.2.0: model choice from the API, model limits in /v1/models, usage on a stream's last chunk; a closed model gives its NPU memory back 11:12 onwx 1.1.0: the KV caches grow with the conversation (27B at 262144: ~16.7 -> ~8.3 GB at load) Oct 21 16 days ago 17:02 onwx 1.0.0: LFM2.5 24B-A2B (converted on your PC, two-block int4, chunked prompts ~430 tok/s) Oct 16 21 days ago 12:26 onwx 0.9.0: Qwen3.6 35B-A3B experts from the SSD (~7 GB of memory), images, faster MTP; changelog shown with updates 10:13 onwx 0.8.0: Qwen3.6 35B-A3B (MoE, int4, all on the NPU: ~20 tok/s, ~22 with MTP; prompts 1K 3.1 s vs FLM 10.2 s), MoE host steps in C 03:12 onwx 0.7.0: Ternary Bonsai 2 9B / 4B (2-bit, images, fast chunked prompts), faster long prompts for llm-jp-4.1 8B and Ternary Bonsai 2 27B (pfa5) Oct 15 22 days ago 13:15 onwx 0.6.2: Ubuntu 26.04 - XRT from AMD's Lemonade PPA (python3-xrt, libxrt-npu2) in the README and the missing-pyxrt message, no build needed 12:40 onwx 0.6.1: faster reading of what agents add after a long context (Hermes run 989 -> 778 s), tool calls written inside the thinking, API max output tokens setting (default 8192), an anchor checkpoint at user messages, clearer XRT diagnosis on Linux 08:02 onwx 0.6.0: polling while reading prompts, opt-in CPU glue for prompt reading, load/unload button, updated FLM comparison 02:35 onwx 0.5.0: faster long-prompt attention, prompts past 32K, long-context fixes, MTP off past 24K, FLM comparison Oct 14 23 days ago 21:38 onwx 0.4.0: all-NPU prefill (1K/2K chunks), int8 KV cache Oct 13 24 days ago 19:56 onwx 0.3.0 01:05 0.2.2: Qwen3.5 4B and llm-jp-4.1 8B thinking in the catalog; prefill fix for models whose layer 0 is attention Oct 12 25 days ago 22:05 0.2.1: Qwen3.5 9B (int4, images; prefill on the NPU) in the catalog, settings text fix 15:06 0.2.0: context up to 262144 tokens, faster long prompts 09:49 Download size in GB (7.6 GB) 09:45 onwx 0.1.0: LLMs entirely on the AMD Ryzen AI NPU (XDNA2) 
