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
Updates: an interruption that a later finished update settled says so in the same red alert (#3578)
e17efa5 Niko1221 committed 28 days ago
Reuse a retained legacy model library through a guarded native data root, without moving weights (#3362)
3f81245 Niko1221 committed 28 days ago
Web page: model settings rows no longer overflow their card on narrow windows (#3601)
9471faa Niko1221 committed 28 days ago
Web page: keep long words and wide tables inside the viewport on narrow screens (#3594)
3b51719 Niko1221 committed 28 days ago
Web page: use an edited API key at once, even when browser storage is blocked (#3435)
2ba95cf Niko1221 committed 28 days ago
Drafting: strata launch --drafting auto|mtp|dflash2|off, one suggestion table, and the DFlash2 drafter fetched pinned and packed on the person's PC
437a606 Niko1221 committed 28 days ago
Merge f-023-df2: DFlash2 drafter (opt-in --draft dflash2) merged onto the combined branch, builds on CUDA, HIP and Metal
f1b6a21 Niko1221 committed 28 days ago
dsv41: --pack-info (the pack as a description: type map, bytes an expert/a token, tiers, tensors with no GPU kernel)
9ee8b0c Niko1221 committed 28 days ago
gfx1151 on Windows: a bare --prefill auto may reach 8192 again (the 6144 cap of #1630 stays for other cards); STRATA_GFX1151_CHUNK=0 keeps 6144
051db74 Niko1221 committed 28 days ago
dsv41: IQ2_XS device kernels (bit-exact vs ggml), Q5_K/Q4_K/Q6_K embedding rows on the host, i-quant expert test, CPU-lane bench reference
49c157f Niko1221 committed 28 days ago
Merge f-023-tune: Work profile CPU workers ceil(0.8 x physical cores), VRAM reserve capped at 1536 MiB on 16 GB cards (picture floor only for a GPU reader); #3493 conversation cache sized automatically after load; #3438 bench reports an engine that did not stop
76ac584 Niko1221 committed 28 days ago
tests: setup no longer tips about the conversation cache (the engine sizes it, #3493)
5de8d26 Niko1221 committed 28 days ago
dsv41: Strata AVX2 i-quant rows opt-in (DS4_GG_IQ256=1) until the engine A/B
20fed55 Niko1221 committed 28 days ago
serve: the conversation cache is sized from the free RAM when --conversation-cache-mib is not given (#3493): 40% of what is left above the parking floor, at most 8192, off under 1024, one log line; an explicit value (0 too) is kept; harness_reuse_probe.py; docs
f0010a6 Niko1221 committed 28 days ago
Merge f-023-tls: #3722 HTTPS uses the OS certificate store (truststore, certifi fallback) in setup, Control, serve and the fetch tools; the app host adds the system roots; a certificate failure says plainly what is intercepting and what to try
ed0ed38 Niko1221 committed 28 days ago
tls: downloads check the server against the operating system's certificates; a failed check gives a plain message (#3722)
c4aec61 Niko1221 committed 28 days ago
A big message on a chat near the compaction threshold summarises first; the real-engine script can also run against a server that is already up
3ae6b30 Niko1221 committed 28 days ago
Summarize and continue asks for the answer instead of sending the refused message a second time; agent README describes the context pieces
ebaeda8 Niko1221 committed 28 days ago
Tests for the two ways to send a message that is too large (file in the chat folder, shortened preview) and the host route
815256d Niko1221 committed 28 days ago
The chat checks a message against the model's window before sending it, handles /compact itself, and explains a full context in plain words
b148623 Niko1221 committed 28 days ago
bench: a failed engine stop is loud and fails the run (#3438): stop_engine returns why the process may still be alive (taskkill result and exit, the 60 s and 30 s timeouts), the run prints an ERROR, writes ENGINE-NOT-STOPPED.txt, exits 6 and reports ok false
f8f4e8a Niko1221 committed 28 days ago
app: the Work profile takes 80% of the physical cores rounded up (the engine's own count at most), keeps no more than 1536 MiB on a card of 16 GB or less, and a CPU picture reader no longer raises the reserve to 2048 (RTX 5070, Gyro-S 64K: 9 workers + 2048 33 tok/s, 5 + 1536 46)
acf63be Niko1221 committed 28 days ago
A marker left by a force-killed app no longer blocks the next load when its process number belongs to a later program
098fd54 Niko1221 committed 28 days ago
A look-only shell command with an & or a bracket inside quotes no longer asks for approval
a7cd2de Niko1221 committed 28 days ago
Compaction reads the prompt size the model server measured, and a summary request always fits the window
be58604 Niko1221 committed 28 days ago