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
Merge f-027-perf: 8,192-token prompt chunk on Strix Halo (11-15% faster prompts)
2b5cadd Niko1221 committed 25 days ago
Strix Halo (gfx1151): an 8,192-token prompt chunk on the unified-memory APU (11-15% faster prompts)
e7e48be Niko1221 committed 25 days ago
arch_defaults_test: the gfx12 table is nine switches now (five prompt, four decode); check the apply path against a clean environment
1d380fa Niko1221 committed 25 days ago
CUDA HC: the wide up read, one warp per row, opt-in (STRATA_GR_UP_WIDE=1)
d18d8ba Niko1221 committed 25 days ago
gfx12: the four bit-identical decode switches back on (ATTN_LANECELL, ATTN_MERGE_V2, TSUM, EXPDEC), measured on p620 R9700 IQ3_S 8K: +3.5% decode, 3/3 pairs, answer md5 identical; docs: the PLE residency policy and its numbers; ple_reader_test: table_bytes
3fd5a74 Niko1221 committed 25 days ago
Merge f-027-strix: DeepSeek V4.1-Flash on Ryzen AI Max (Strix Halo, gfx1151) on Windows
3f26ca5 Niko1221 committed 25 days ago
setup: accept --feed (the DeepSeek engine's package comes from it)
a6543d6 Niko1221 committed 25 days ago
Strix Halo: the DeepSeek engine follows --feed, and the unified plan counts the BIOS carve-out
932bda5 Niko1221 committed 25 days ago
Engine: keep the PLE table resident in RAM when the PC can afford it (opt-out STRATA_PLE_AUTO_RAM=0); measured on p620 R9700 IQ3_S: 8K prompt PLE 845 -> 132 ms, 2125 -> 2611 tok/s, answer md5 identical
3b34b6c Niko1221 committed 25 days ago
Strix Halo (gfx1151) on Windows: DeepSeek V4.1-Flash support and the 0.2.6 shared-memory fix
6c70d4b Niko1221 committed 25 days ago
WIP not-default: DS4_NO_WAIT_SPLIT (no commit at the routing wait; A/B unfinished) and DS4_PHASES (decode host phase timers)
1b667c7 Niko1221 committed 25 days ago
DS4_PLACE_HASH: one hash of which experts hold a VRAM slot (the placement changes with the saved profile)
4acede6 Niko1221 committed 25 days ago
Merge main (41cf05a06: README model order, docs only) into f-026-rel
53548c8 Niko1221 committed 25 days ago
README (all 7 languages): models in the order Qwen3.8-Flash-Next, Qwen3.6-35B-A3B, DeepSeek V4.1-Flash
41cf05a Niko1221 committed 25 days ago
Merge main (ed4b4db40: README, docs and icons) into f-026-rel; AGENTS.md keeps main's text with Version 0.2.6, engine 0.1.51; the installer names in the READMEs and docs are the 0.2.6 ones
4f0f933 Niko1221 committed 25 days ago
README translations (zh-CN, ja, de, fr, es, pt-BR) rebuilt in the structure of the new README; the README links source setup to SOURCE_INSTALL.md and names the three 0.2.5 installers
af8c6f8 Niko1221 committed 25 days ago
AGENTS.md and AI_SETUP.md: lead an agent to the installer, install.sh/install.ps1 and strata launch --plan then --yes --no-tui; source setup only as the fallback; provider order ISTA-DASLab, Unsloth, Gyro, others
e99c4d2 Niko1221 committed 25 days ago
docs: add the guide index, the source-install page, the AI reference and the media kit (the owner's pages, updated to 0.2.5)
b8b4ca0 Niko1221 committed 25 days ago
Host test: the source check for 'a version that is not downloaded starts with Read pictures on' accepts the braced else that the DeepSeek 8K context fix introduced (test only)
79ec8e2 Niko1221 committed 25 days ago
Version 0.2.6: Terminal and app 0.2.6, engine and Control 0.1.51 (every Qwen and DeepSeek package, Metal included)
27a8ed9 Niko1221 committed 25 days ago