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++

setup: a tip for Gyro installs on the ready-made engine (the fork's prompt kernels need a source build)

Niko1221375f76fef032
diff --git a/setup.py b/setup.py
index 811921f99..4c1c92b62 100644
--- a/setup.py
+++ b/setup.py
@@ -6601,6 +6601,10 @@ def main() -> int:
ok(f"size: {model}")
if MODELS[model].get("license"): # the ready-made packs: each size has its own license
say(f" Its license: {MODELS[model]['license']}")
+ if family == "gyro": # a tip, nothing changes (recommend, never force)
+ say(" Tip: Gyro's prompt kernels come from agentionai's llama.cpp fork. The engine setup installs reads prompts "
+ "through a slower path (measured 20-30% slower on an AMD R9700; answers speed is the same); compile with "
+ "-DSTRATA_GGML_DIR=<agentionai/llama.cpp> for the fast one (docs/GYRO.md).")
tag = fam["tag"] + model # names of the pack, config and start script
small = min(x["vram_gb"] for x in chosen) # each card keeps its layers' KV of the whole context
rec_ctx = 32768 if small < 14 else 65536 if small < 20 else 131072