onw 0.14.1: memory-saving mode swaps the 16-token and 1-token graphs (one set in memory at a time)

onw 0.14.0: memory-saving mode (16-token graphs only) for many-part models; no auto-reload after updates when auto-load is off

onw 0.13.2: NPUW auto-off for many-part models (Granite, llm-jp); NPU driver resource errors not recorded as route failures

onw 0.13.1: NPUW weight sharing for Qwen3.5 16-token graphs (no boolean mask constants)

onw 0.13.0: NeoHorse-1-4B, Granite 4.2 3B / 8B, llm-jp-4.1-8b-thinking; granite / llama converter; Panther Lake listed

onw 0.12.1: load progress, download progress for updates, stale manager.json fix

onw 0.12.0: Ornith-1.5-9B, earlier turns' reasoning_content to the template, --deltanet-fmt q8g128

onw 0.11.3: Ubuntu launcher no longer crashes (namespace shadowing guard, console-script launchers)

onw 0.11.2: 16 GB PCs no longer mark models as SSD-streamed

onw 0.11.1: installs again (OpenVINO pre-release index), code review fixes, gpt-oss INT8 embedding

onw 0.11.0: OpenAI gpt-oss-20b (gpt_oss: MoE, attention sinks, YaRN, harmony), INT8 code overflow fixes

onw 0.10.5: the compile cache prunes itself after a load; attention graphs in onw blob store

onw 0.10.4: keep the .bin mapped for the NPUW weights bank (crash at first inference); REP pipeline for the split parts

onw 0.10.3: NPUW works again for the split parts (Lunar Lake memory + compile time); 3720 DCOFF crash fix

onw 0.10.2: fix the experimental switches flipping back off

onw 0.10.1: experimental NPUW settings (MoE, FUNCALL_ASYNC, weightless cache); less CPU in long chats

onw 0.10.0: long contexts with a paged INT8 KV cache (32K default, up to the model max); Gemma RoPE fix; Ubuntu launcher fix

onw 0.9.2: tool calls (function calling) in the OpenAI-compatible API

onw 0.9.1: the window closes with the tray; 'Check now' always answers; parallel update downloads

onw 0.9.0: LFM2.5 support (convert it yourself), LFM models out of the list, torch-free convert

README reorganized (install first); technical notes moved to TECHNICAL.md

onw 0.8.5: after an update the onw window reopens on the new version

onw 0.8.4: English / Simplified Chinese UI, new icon, README banner

README: screenshots of the onw window

v0.8.3: update notice as a bottom-right card with progress; open windows reconnect to the restarted tray

v0.8.2: predicted prefetch while decoding off by default (CPU cost on 8-core Lunar Lake); prompt prefetch stays

v0.8.1: auto-update (check at start and daily, tray notice, one-button update + restart + model reload), CHANGELOG, clearer auto-load setting

v0.8.0: faster SSD streaming without more memory - prompt prefetch into cache slots, predicted next-layer expert prefetch while decoding (helper thread, separate read queue), single-copy reads

v0.7.2: Linux imports stored blobs through a stream (a mapped blob failed on the NPU 4000 driver); ONW_OWN_BLOBS=0 / ONW_IMPORT switches

v0.7.1: model list notes and catalog memory updated for onw 0.7

v0.7.0: 20-60% less memory (own blob store, vision loaded on demand and freed, host tables mapped, direct-I/O expert load); REAP streams too; first-on-Intel-NPU SSD streaming note

v0.6.1: layer-major prompt processing when experts stream from the SSD (each layer's experts read once per prompt)

v0.6.0: stream MoE experts from the SSD (LRU cache, direct I/O) when the model does not fit; expert memory setting

README: to our knowledge the first public all-on-NPU VLM / Gated DeltaNet / MoE (survey Sept 2026)

v0.5.0: restore setup.sh / start.sh (they were uploaded empty)

v0.5.0: installers retry each download and normalise line endings

v0.5.0: install.sh without a stray CR

v0.5.0: setup.sh with LF line endings again; install.sh strips CR before running setup.sh

v0.5.0: setup no longer upgrades pip (a new pip failed to install packages with scripts), rebuilds the venv once on failure, installs the web view package on upgrade

v0.5.0: new window (HTML in a web view; compact table, state colours, narrow layout), manager in the tray (downloads survive closing the window), pywebview

v0.4.10: compile probes out of process (compiler crashes survive), NPUW config aligned with the standalone Lunar Lake ports, compiler version in logs

v0.4.9: version in titles and logs, compile fallbacks (NPUW -> 1 tile -> plain -> drop block size) remembered per driver

v0.4.8: NPUW fixes (no SLICE_OUT, bool constants), NaN check, ONW_SIZES, Linux driver 1.38 note, diagnostic tools

v0.4.7: no NPUW for MoE models, bounded hold on broken characters, clear-cache button, save button on top

v0.4.6: NPUW on NPU 4000+ uses compiler dequantization (same baseline as OpenVINO's NPUW LLM mode)

v0.4.5: NPUW weight sharing (auto on NPU 4000+, experimental), taller list rows on Ubuntu

v0.4.4: load/unload UI with model locations, single instance, background polling, per-PC memory estimate

v0.4.3: choose the model save folder and add folders that already hold models

v0.4.2: one-line installers (install.ps1 / install.sh), beginner-friendly getting started

onw 0.4.1: Japanese README (README.md) + English (README_en.md); setup/start call find_python.bat by path