Skip to content

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++

FP4-Local engine floor 0.1.45 (the split release's engine has the NVFP4 n-gram table)

Niko1221e7f7dad3e0de
diff --git a/setup.py b/setup.py
index d7b8a12be..11e299c8f 100644
--- a/setup.py
+++ b/setup.py
@@ -296,7 +296,7 @@ FP4LOCAL_PREPARED = "prepared" # <models>/nvidia--Qwen3.8-
# the first release engine with NVFP4 experts (a source build of this checkout has them whatever its number says)
NVFP4_ENGINE = (0, 1, 44)
# the first engine that reads NVIDIA's NVFP4 n-gram table (strata.ple.format = nvfp4x16); a source build of this checkout has it
-FP4LOCAL_ENGINE = (0, 1, 46)
+FP4LOCAL_ENGINE = (0, 1, 45)
# agentionai's Gyro files (Qwen3.8-Flash-Next with trellis-coded, Hadamard-rotated experts, TQ2_T / TQK6 / TQK7): one
# GGUF per size, the n-gram table inside it (hence the 58 / 92 GB downloads for 27.6 / 35 GB of memory). The engine