The most intelligence a laptop can fit
A local LLM server for AMD Strix Halo, native on Windows
Halowin Bubo runs Qwen3.8 Flash Next on your laptop. The new app helps you download the model, starts it and shows what it uses. Version 0.1.2 runs the ISTA GSQ-RCO files IQ3_S, IQ3_XXS, IQ2_XS and the Flash Next Coder as published. It adds an optional fast mode for prompts of 4,096 or more new tokens, a smaller context cache and in-engine abliteration, off by default.
OpenAI-compatible API · tool calling · up to 262K context · tested with the Pi coding agent
Needs Windows 11 and an AMD Strix Halo machine. Qwen3.8 Flash Next needs 128 GB of memory and about 113 GiB of free disk. Smaller models need less. On a 96 GB machine, use a lower-memory model: the lower-memory build, an ISTA file or the Coder. MIT license. See the known issues.
Qwen3.8 Flash Next: 125B parameters (6B active) at about 70 W
66.2 tok/s
writing, in a long agent session at about 70 W
1,639 tok/s
reading a 32K prompt, with fast mode
Faster on the same laptop
| Workload | Bubo 0.1.2 | Other engines, tok/s | |
|---|---|---|---|
| gufo Windows 0.24.0 build b72e01b | gufo-win 0.5.0 release v0.5.0 | ||
| ProseWriting speed (decode). tok/s | 57.1 | 45.3 | 53.4 |
| Agent sessionWriting speed (decode). Long session, thinking on. tok/s | 66.2 | 48.0 | 54.7 |
| Code editsWriting speed (decode). Edits that reuse your file’s text. tok/s | 109.0 | 44.0 | 110.6 |
| Reading a 32K promptReading speed (prefill). tok/s | 1,639* | 1,413 | 1,444 |
*Experimental fast mode for long prompts, off by default. Turn it on with --prefill-mode fast (in CLI), or with Prefill mode (in GUI).
Same laptop at about 70 W, same model file (Huihui UD-Q4_K_XL) for every engine. Method and every row
Smaller files that lose very little reasoning
Bubo also runs ISTA’s lower-memory files of the same model. They leave room for more sessions at longer context, a smaller GPU carve-out or other programs beside the model.
Start in four steps
start.cmd