Qwen3.8-4B-Empero-Q6_K-llamafile

Qwen3.8-4B Empero Q6_K llamafile

This is an unofficial, local, self-contained Empero-named llamafile package of the already-quantized Qwen3.8-4B-Q6_K.gguf published by Empero AI. Empero AI is credited here as the developer/publisher of the distilled empero-ai/Qwen3.8-4B model and the publisher of its direct quantized GGUF repository. This package combines the pinned llamafile 0.10.5 runtime, the existing Q6_K GGUF bytes, and packaged command-line defaults in one executable APE/ZIP file.

This package only verified and embedded the exact existing Q6_K source GGUF bytes. It did not itself train, fine-tune, merge, convert, edit, or re-quantize the model. This file is executable code plus model data, not merely a weights file; verify its SHA-256 before running it.

Publisher claims:

Empero describes Qwen3.8-4B as a 4B, full-parameter supervised fine-tune with off-policy distillation using approximately 45,000 curated teacher traces.

Empero describes it as the text path of a vision-language-model base, says the fine-tune was text-only, reports a native context of 262,144 tokens, and says inherited vision was not evaluated.

Those training, benchmark, context, capability, and vision statements are attributed publisher claims. They were not independently validated while building this package. No inference or benchmark was run as part of this packaging task.

Exact lineage and roles

The lineage is deliberately explicit:

Hugging Face model tree: this Empero-named llamafile → empero-ai/Qwen3.8-4B-GGUF (direct quantized GGUF source) → empero-ai/Qwen3.8-4B (Empero full-parameter fine-tune/distill) → Qwen/Qwen3.5-4B (base).

This package is one derivative level below the Empero GGUF in the Hugging Face model tree: the direct parent is Empero AI's GGUF repository, while the executable merely embeds its existing Q6_K bytes.

1. This package — Qwen3.8-4B-Empero-Q6_K.llamafile; an APE runtime plus ZIP payload assembled locally.
2. Exact GGUF source — empero-ai/Qwen3.8-4B-GGUF, published by Empero AI, revision 391fc7d103e3942a408def3e4f51c2f85d464417, file Qwen3.8-4B-Q6_K.gguf.
3. Direct model repository — empero-ai/Qwen3.8-4B, developed and published by Empero AI, revision c83cb7aa2999d2f35c43e9ae0634a30eb8985a1e; this is where Empero describes its text-only fine-tuning/distillation work.
4. Original base — Qwen/Qwen3.5-4B, a Qwen3.5 vision-language causal model base.

The model-card base_model points to the direct GGUF source because this derivative packages that repository's exact quantized artifact. base_model_relation: quantized describes the relationship of the direct parent; it does not claim that this llamafile build performed quantization.

Model and package facts

Artifact identity

Always compare the complete 64-character digest before execution:

shasum -a 256 Qwen3.8-4B-Empero-Q6_K.llamafile

The result must be:

4f9fc8a4fef4b56bb427dcf1a893a813636cac3ca54f1797a35ae7bdb47a7941  Qwen3.8-4B-Empero-Q6_K.llamafile

Exact components and retained tools

The source GGUF and .args are ZIP-stored rather than compressed and their data offsets are aligned to 65,536-byte boundaries for mapping. The build manifest records the exact local-header and data offsets. Runtime ZIP members remain part of the pinned upstream executable.

Exact packaged defaults

The external .args file is byte-identical to the embedded .args. It contains one item per line:

-m
/zip/Qwen3.8-4B-Q6_K.gguf
--ctx-size
8192
--jinja
--temp
0.6
--top-p
0.95
--top-k
20
--min-p
0.0
...

These defaults mean:

• /zip/Qwen3.8-4B-Q6_K.gguf selects the embedded weights;
• --ctx-size 8192 chooses an intentionally conservative 8,192-token context;
• --jinja enables the embedded model chat template path;
• sampling defaults are temperature 0.6, top-p 0.95, top-k 20, and min-p 0.0;
• the final ... passes user-supplied arguments after the packaged defaults.

To request a different context, append an override such as --ctx-size 16384. Larger values can sharply increase memory use. An override cannot prove quality or stability at that context and cannot exceed what the model/runtime/build actually support.

Native context versus packaged context

Empero and the Qwen base documentation report 262,144 tokens natively. This package intentionally defaults to 8,192 tokens to reduce the initial KV-cache/RAM burden on ordinary local machines. The smaller default does not modify the model weights or reduce the publisher-reported architectural maximum; it changes only the launch default. Conversely, the publisher-reported native context is not a promise that this Q6_K llamafile will fit in memory or perform well at 262,144 tokens on a given computer.

Use from a terminal

No execution was performed during this build. The following commands are instructions for a later, user-controlled test.

On macOS or Linux, make the artifact executable if the copied filesystem did not preserve its mode:

chmod +x Qwen3.8-4B-Empero-Q6_K.llamafile

Start the terminal/chat path with packaged defaults:

./Qwen3.8-4B-Empero-Q6_K.llamafile

Run a one-shot prompt with a bounded output length:

./Qwen3.8-4B-Empero-Q6_K.llamafile -p 'Write one sentence about local inference.' -n 128

Override the packaged context for a machine with enough memory:

./Qwen3.8-4B-Empero-Q6_K.llamafile --ctx-size 16384

Local browser UI and API

Start the built-in server bound to loopback only:

./Qwen3.8-4B-Empero-Q6_K.llamafile --server --host 127.0.0.1 --port 8080

Then use:

• browser UI: http://127.0.0.1:8080/
• OpenAI-compatible API base: http://127.0.0.1:8080/v1

Example API request after the server is started:

curl http://127.0.0.1:8080/v1/chat/completions \
  -H 'Content-Type: application/json' \
  -d '{"model":"local","messages":[{"role":"user","content":"Say hello."}],"max_tokens":64}'

Client details can vary with the pinned runtime. Inspect the local runtime's help in a later authorized execution session if an option is rejected; this build deliberately did not execute either the runtime or the final model.

Localhost and executable security

127.0.0.1 restricts the example server to the local host. Binding to 0.0.0.0, a LAN address, or an Internet-facing interface can expose prompts, model outputs, and an unauthenticated service to other machines. Use a firewall and a properly authenticated reverse proxy before any deliberate network exposure.

A .llamafile is executable software. Verify the exact SHA-256, obtain it from a trusted source, retain normal operating-system protections, and do not run it with elevated privileges. Hosted inference is disabled for this repository metadata because this artifact is a standalone executable rather than a normal hosted-inference checkpoint.

Portability, platform, storage, and RAM

The single .llamafile can be copied to a USB drive or portable SSD. The destination filesystem must support a file of 3,913,892,851 bytes and, for direct Unix-style launch, executable permissions. FAT32 cannot store files of 4 GiB or larger; this particular file is below 4 GiB, but free space, filesystem limits, mount options, and executable-bit behavior still matter. Copy the checksum ledger with the artifact and verify after transport.

This package is intended primarily for supported 64-bit macOS and Linux environments covered by llamafile 0.10.5. CPU architecture, OS policy, available acceleration, filesystem execution policy, and runtime compatibility can all affect whether it starts. The Q6_K weights alone occupy about 3.318 GiB; runtime state and KV cache require additional RAM. An 8 GB machine may be tight; 16 GB or more is a more practical starting point, especially if context is increased. This is planning guidance, not a measured requirement, because the artifact was not executed.

The artifact is 3,913,892,851 bytes, which is below 4 GiB and therefore below the commonly documented direct-Windows APE size boundary. Windows execution was not tested for this build; size eligibility is not an execution result. On Windows, an .exe suffix or other platform-specific handling may be needed. Do not infer compatibility from size alone.

Text-only package and projector status

The original Qwen3.5-4B base is a vision-language model, and Empero describes Qwen3.8-4B as its text path. Empero says its fine-tune was text-only and that inherited vision was not evaluated. The direct GGUF repository at the pinned revision provides no projector for this package.

Accordingly, this llamafile does not include or enable a projector. There is no mmproj file and no --mmproj argument. It is a text-generation package even if some vision-related structure is inherited in the broader base lineage. Supplying an image to a client will not create a supported vision path without the correct processor/projector pairing, runtime support, and independent validation.

Intended uses

• local, offline text generation on a compatible personal computer;
• portable one-file experimentation from internal storage, USB, or a portable SSD;
• terminal chat, bounded one-shot prompting, or a loopback-only local API/UI;
• reproducible inspection of the exact Q6_K weights and pinned runtime.

This package is not a hosted service, not an independently benchmarked release, not a vision-enabled package, and not evidence that the publisher's training or capability claims reproduce in llamafile 0.10.5.

Limitations and caveats

• No model execution, inference, response-quality test, benchmark, or performance test was performed during packaging.
• Publisher training, benchmark, context, capability, language, and inherited-vision claims were not independently validated.
• Q6_K is lossy quantization; behavior can differ from higher-precision weights.
• The 8,192-token default is conservative and much smaller than the publisher-reported native context.
• Long contexts increase RAM use and may be slow or fail to allocate.
• Runtime behavior is pinned to llamafile 0.10.5; newer or older runtimes may differ.
• This package has no projector and supports text generation only.
• Windows execution was not tested, even though the artifact is below the >4 GiB direct-Windows boundary relevant to larger APE files.
• Model outputs can be inaccurate, biased, unsafe, or unsuitable for high-stakes decisions. Apply human review appropriate to the use case.

Build and archive verification

The retained build-Qwen3.8-4B-Empero-Q6_K-llamafile.py script performs a no-clobber build, pins source/tool sizes and hashes, isolates HOME/cache/TMPDIR under an owned temporary subtree, calls the matching hash-pinned zipalign-0.10.5 through /bin/zsh with -j0, and removes temporary state on success or failure.

Verification covered:

• source and shared-tool regular-file/no-symlink admission;
• source, runtime, and zipalign exact byte size and SHA-256;
• ZIP CRC/integrity and unique archive member names;
• exact embedded model and .args names, sizes, hashes, and ZIP_STORED method;
• 65,536-byte data alignment;
• byte-identical retained and embedded .args;
• unchanged source stat/hash and source-directory root inventory after packaging;
• unchanged shared-tool files and shared-directory root inventory;
• exact final root allowlist and checksum-ledger verification;
• Hugging Face card metadata parsing/validation as recorded in BUILD-MANIFEST.md.

An execution/inference check was intentionally deferred to the parent workflow and is not claimed here. The build script did not upload or publish any file; a later private Hub publication, if present, is a separately authorized post-build action.

License and attribution

The model/source lineage is distributed under Apache-2.0 according to the supplied source metadata. The complete license text is retained in LICENSE. The llamafile runtime and packager are pinned third-party software components; retain their upstream copyright/license notices and review the upstream project when redistributing them separately.

Upstream repositories:

• GGUF: https://huggingface.co/empero-ai/Qwen3.8-4B-GGUF
• direct model: https://huggingface.co/empero-ai/Qwen3.8-4B
• original base: https://huggingface.co/Qwen/Qwen3.5-4B
• llamafile: https://github.com/Mozilla-Ocho/llamafile