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
MCP chat: at the tool-round limit the model gets one final reply with no tools from the results already found, instead of the turn ending without an answer (PR #3741 by i-Vladimir-i)
e0469a5 i-Vladimir-i committed 24 days ago
docker: the build context carries the nested tools/ scripts the Terminal bundle reads (#3974)
d398f35 天空Blond committed 24 days ago
procs: an engine adopted by systemd --user is found as an orphan, and ending it waits on the start token until the process is really gone (PR #3254 by toel1234)
dc1fd36 toel1234 committed 24 days ago
App sharing: the shared and tunnel address answers a browser's CORS preflight and sends Access-Control headers when an API key is set; without a key nothing changes (PR #3853 by MaKla89, #3208)
eb1c400 MaKla89 committed 24 days ago
Web page: a request that timed out in the queue shows a Timeout badge in Recent requests (PR #3991 by thisisspartan)
cd72acc thisisspartan committed 24 days ago
Merge f-027-cuda (CUDA speed work: wide HC up read on sm_120, residency-biased routing on small cards, STRATA_TRACE, release-notes draft)
4abccbb Niko1221 committed 24 days ago
Merge f-027-win (main 2d07d1c0b with the DeepSeek slab scoping 32 MB unified / 8 MB discrete plus automatic growth, the dq kernel arch choice, and the HIP-only guard for the gfx11 WMMA gemm call)
f92cfa4 Niko1221 committed 24 days ago
serve: the lifeline identity is the child's start time, the sampler survives a bad source (#3590, #3407)
b3db5a1 justindressler committed 24 days ago
App: the source-checkout server is started without an empty STRATA_API_KEY (#3851); a card taken out of the PC no longer decides the engine make, and a PC with two makes accepts a build for either (#3892)
9a1e768 Niko1221 committed 24 days ago
Monitor: the host reads the graphics card every 3 s only while a model runs or the Monitor is open, every 30 s otherwise, and the PCIe sample only while a model runs (#4066)
085dcde Niko1221 committed 24 days ago
serve: accept Codex spellings of a namespace tool (#3972)
fd8d37e Righteous-AI committed 24 days ago
Chat: the transcript follows the answer only while the reader has not left; a wheel notch, finger or key moving up lets go at once, in the app chat and the served web page (#4014)
89a68d8 Niko1221 committed 24 days ago
conversation cache: never park one conversation with another's retained K/V (#1503)
f8ff342 Stefanos Anastasiou committed 24 days ago
Manual mode: the look-only shell allowlist takes only options known to look (#3975)
f681b9a Niko1221 committed 24 days ago
qwen35: a penalty for a history token outside the logit row no longer indexes past the row (#3911)
6c01f85 Niko1221 committed 24 days ago
Merge f-027-kva (docs/KVA.md: the shipped opt-in's numbers on the R9700 and the P100, and the decode trade on fast GPUs)
3a604e0 Niko1221 committed 24 days ago
docs/KVA.md: the R9700 and P100 numbers for the shipped opt-in (split 36), and the trade stated plainly
6f45b24 Niko1221 committed 24 days ago
DeepSeek engine: the RDNA3 WMMA dense GEMM call in gemm() is HIP-only (wmma_gemm.cu is part of the AMD builds; main's 0eea4a3d5 called strata_pf_gemm_f16_ld unconditionally, so a CUDA build failed to link: unresolved external)
063b3c5 Niko1221 committed 24 days ago
ds4_iq.cu: the NUL character literal in dq_use_old_kernel had been written as a raw NUL byte by the merge script (nvcc: quoted string should contain at least one character); it is a backslash-zero escape again
20c2f42 Niko1221 committed 24 days ago
prompt: the RAM store's post-prompt preload is off by default - it costs a 4K prompt 4.7 s and the decode that follows gains nothing
259532b Niko1221 committed 24 days ago
KVA, when switched on, approximates layers 36-47 with a 2048-token exact tail (the measured configuration); STRATA_KVA_SPLIT / STRATA_KVA_SUFFIX override
bda3005 Niko1221 committed 24 days ago
Merge f-027-kvaoff (KVA is off by default: the app's and the Terminal's 'Faster long prompts' switch turns it on, shown for Qwen3.8-Flash-Next models only)
aeb70de Niko1221 committed 24 days ago
KVA: the one-time note says how many positions were approximated, not where the tail starts (chunks round the frontier down)
76cd7d7 Niko1221 committed 24 days ago
KVA: keep the geometry gate - a model whose shape KVA was not measured on reads exactly; STRATA_KVA_ANY_MODEL=1 is the research override
25a62e0 Niko1221 committed 24 days ago
Faster long prompts is shown for Qwen3.8-Flash-Next models only (app and Terminal)
72b7696 Niko1221 committed 24 days ago