Qwen3.8-Flash-Next on any consumer hardware: one-click install for Windows / Linux. Strata inference engine, OpenAI/Anthropic API on localhost, optional image input.
C++
Commits on main
ui: Quit Strata in the menu in browser mode (confirm, stopped screen), 8 languages; host: open the page without changing the mode
a8fdceb Niko1221 committed 1 month ago
desktop (Linux): the shell owns the host, the window is its own process; a crashed window moves Strata to the browser (--browser, STRATA_UI=browser, --quit, session file)
8be9e09 Niko1221 committed 1 month ago
host: window mode vs browser mode (app-ui.mjs): Quit Strata, the shell's UI route, app updater unavailable in browser mode with a plain message
e01f766 Niko1221 committed 1 month ago
ui: no route cross-fade (View Transitions) on WebKitGTK, which crashes its UI process on startViewTransition (null pointer in AcceleratedBackingStore::update)
401169c Niko1221 committed 1 month ago
docs: the 12.0 defaults with the final A/B (decode +14-16%, prompts +3-10%), PF_FUSED quality, the ple-io rule
b829a31 Niko1221 committed 1 month ago
setup: --ple-io ram on an SSD too when the RAM can spare the n-gram table (Linux)
b25b309 Niko1221 committed 1 month ago
prefill: fused int8 prompt experts by default on sm_120 (IQ2_S/IQ3_XXS/IQ3_S/IQ4_XS)
959e6c3 Niko1221 committed 1 month ago
experts: gate/up, SwiGLU and q8_1 in one launch (bitwise; default on sm_120)
75afe3f Niko1221 committed 1 month ago
launch: shorter model-list wording (too big, SSD N GB) so it fits 110 columns; docs and previews for the classic-console look and several cards
fa726b4 Niko1221 committed 1 month ago
monitor: the classic Windows console switches to Cascadia Mono when installed, else a solid-block look with palette colours; several cards and CPUs in the Overview (totals, a Cards strip, sockets); whole card names in view 5
1d03976 Niko1221 committed 1 month ago
sm_120: the hyper-connection down projection with one token a block row (bitwise)
9410417 Niko1221 committed 1 month ago
mmvq_il: the sm_120 2-4 column table only on 80+ SM cards; docs
4499712 Niko1221 committed 1 month ago
Docker: smoke-test.sh (the checks run on real cards), executable bits
ae31e3c Niko1221 committed 1 month ago
launch: the card or cards chosen this time go into every launch's own config (--gpu, the Cards step, or the fit rule), so an installed model never keeps its old card and nothing is reinstalled
7bdc1ff Niko1221 committed 1 month ago
Docker: GPU/GPUS take effect on a volume that already has the model, tuning wording, build id of the source build
a00ed93 Niko1221 committed 1 month ago
Dockerfile (source): no NCCL in the picture reader, so the package check passes
404713f Niko1221 committed 1 month ago
Docker: keep only gguf-py and the ggml marker of the baked llama.cpp; root Dockerfile PATH fix
4895b1c Niko1221 committed 1 month ago
desktop: cards-e2e.mjs, the Graphics cards choice end to end through the host API
039d2d4 Niko1221 committed 1 month ago
app: the Graphics cards choice on the model page (the recommended option pre-selected, neutral style), passed on install and load
33d37dd Niko1221 committed 1 month ago
sm_120: four bitwise decode switches on by default on a 12.0 card
ba2c2d5 Niko1221 committed 1 month ago
mmvq_il: 5-8 column interleaved kernels and the sm_120 tables (bitwise)
75fabbd Niko1221 committed 1 month ago
Docker: pinned llama.cpp baked into the image (no GitHub download on first start), file consistency tests
390336f Niko1221 committed 1 month ago
Docker: launch gets SIGINT back so docker stop ends a download in 1 s; INSTALL.md points to DOCKER.md
7e490ef Niko1221 committed 1 month ago
Docker: DOCKER.md, MODEL_PATH in compose, device check for AMD and Intel
3641daf Niko1221 committed 1 month ago