Qwen3.8-27B UD-Q5_K_XL llamafile

An unofficial, reproducible llamafile packaging of Unsloth's Qwen3.8-27B UD-Q5_K_XL GGUF. It embeds the main model, its FP16 vision projector, and the llamafile 0.10.5 runtime in one file; it is not a new training run or a re-quantization.

Model-tree lineage: this llamafile → unsloth/Qwen3.8-27B-GGUF → Qwen/Qwen3.8-27B. The direct parent is the Unsloth GGUF repository whose already-quantized UD-Q5_K_XL weights are embedded here unchanged.

Model details

Artifact

A llamafile is executable code plus model data, not merely a weights file. Verify the checksum before running a downloaded copy.

What is packaged

• Qwen3.8-27B-UD-Q5_K_XL.gguf
• mmproj-F16.gguf
• llamafile 0.10.5
• Embedded launch arguments pairing the model and projector

The GGUF and projector are stored uncompressed with zipalign -j0. Neither was modified or re-quantized during packaging.

Embedded defaults

-m
/zip/Qwen3.8-27B-UD-Q5_K_XL.gguf
--mmproj
/zip/mmproj-F16.gguf
--ctx-size
8192
--jinja
--temp
1.0
--top-p
0.95
--top-k
20
--min-p
0.0
--presence-penalty
0.0
...

The final ... allows additional command-line arguments supplied by the user.

Provenance and checksums

unzip -t completed successfully against the final artifact.

Run on macOS or Linux

chmod +x Qwen3.8-27B-UD-Q5_K_XL.llamafile
./Qwen3.8-27B-UD-Q5_K_XL.llamafile

The default launch provides both a terminal chat prompt and a local Web UI/API server.

Browser chat UI

Keep the terminal process running, then open:

http://127.0.0.1:8080

127.0.0.1 keeps the server accessible only from the local machine.

One prompt/response

./Qwen3.8-27B-UD-Q5_K_XL.llamafile \
  --cli -p 'Explain the purpose of a llamafile.' -n 256

Server only

./Qwen3.8-27B-UD-Q5_K_XL.llamafile \
  --server --host 127.0.0.1 --port 8080

The server exposes a local Web UI and OpenAI-compatible endpoints under http://127.0.0.1:8080/v1.

Intended use

This package is intended for people who want a single downloadable executable containing:

• a capable local Qwen3.8 27B model;
• the matching vision projector;
• a terminal chat interface;
• a browser chat interface; and
• a localhost API server.

It is particularly convenient for offline or self-contained macOS/Linux use where installing a separate inference runtime is undesirable. Because the model, projector, runtime, terminal chat, browser UI, and local API are bundled into one file, you can keep the llamafile on a sufficiently large USB drive or portable SSD and bring your local intelligence with you. On a compatible machine, that single file is all you need.

Notes

• This is an unofficial packaging project, not an official Qwen, Unsloth, or Mozilla release.
• No training, fine-tuning, alignment changes, or new evaluation results are claimed here.
• Model capabilities, limitations, and safety characteristics are inherited from the upstream Qwen and Unsloth artifacts.
• The 8,192-token packaged default is intentionally below the upstream maximum. Override it by appending --ctx-size N, for example --ctx-size 32768; larger contexts require more RAM and may reduce performance.
• This artifact is already multimodal: the matching mmproj-F16.gguf projector is embedded and selected automatically by the packaged defaults. No external projector file is required.
• Hugging Face hosted inference is disabled because this repository distributes a standalone executable rather than a standard hosted checkpoint.

Windows limitation

This artifact exceeds Windows' 4 GB executable limit and will not run as a Windows .exe.

License and attribution

The upstream Unsloth model card declares Apache-2.0. The embedded llamafile runtime is also Apache-2.0. Retain upstream notices and attribution when redistributing.