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++
tests: the 1.75 bpw set also runs with --tp 1 on 4 AMD cards; release notes: set up again on 4+ AMD cards
Niko12218ac2a8385e51
diff --git a/docs/release-notes-0.2.4.md b/docs/release-notes-0.2.4.mdindex 31581f821..5428c83a3 100644--- a/docs/release-notes-0.2.4.md+++ b/docs/release-notes-0.2.4.md@@ -56,6 +56,7 @@ These are the 2.00-bit set. The 2.25-bit set reads 13% more data per token, so e- Not in this release: Mac and Strix Halo. DeepSeek V4.1-Flash does not run there yet.- Output on 12 GB cards is 4 to 8 tokens/s. That is usable for chat and slow for long agent runs.- The same prompt can give a slightly different answer from one run to the next. At some words the model's top choices are nearly tied, and small differences in how the numbers are added up decide between them.+- If you set up DeepSeek on 4 or more AMD cards before 0.2.4, run setup again to get the faster prompt settings.## Updatingdiff --git a/tools/test_setup_dsv41.py b/tools/test_setup_dsv41.pyindex d62e217d2..358aacc77 100644--- a/tools/test_setup_dsv41.py+++ b/tools/test_setup_dsv41.py@@ -86,7 +86,7 @@ class Fit(unittest.TestCase):def test_four_amd_cards_run_the_ista_sets_with_a_plain_layer_split(self):# r9700, 2026-10-23: 8K prompt 370 tokens a second with --tp 1 against 196 with the engine's default (2)- for model in ("IST-DAS-2.00bpw", "IST-DAS-2.25bpw"):+ for model in ("IST-DAS-1.75bpw", "IST-DAS-2.00bpw", "IST-DAS-2.25bpw"):p = dsv41_fit.plan(FOUR, 251.0, 900.0, model=model)self.assertEqual((p["tp"], p["args"]), (1, ["--tp", "1"]), model)# two and three cards keep pairs; other makes keep pairs; one card needs no flag; the Unsloth set is not changed