Qwen3.8-27B for Splash · experimental weight-edit conversion

A community-built, 4-bit Qwen3.8-27B package for local use with Splash on Apple silicon. It includes the target model, DFlash 2 draft, vision encoder, and tokenizer.

This build applies weight differences from huihui-ai's abliterated model to Inco AI's Splash package. Abliteration aims to reduce learned refusal behavior. It can also affect reasoning, accuracy, and appropriate caution. This conversion does not establish a refusal-reduction rate or preserve all of the original model's capabilities.

Use it for supervised local experiments, writing, and model-behavior research. For a general-purpose starting point, consider the original Splash package. This is an independent derivative; the upstream authors have not endorsed it.

Run locally

The upstream package specifies Apple M3 or newer, macOS 26.4 or later, and at least 36 GB of unified memory, with 48 GB or more recommended. The checks below used an M5 Max with 128 GB; they do not validate the minimum configuration.

brew install incoai/tap/splash
splash serve --model audreyt/Qwen3.8-27B-Splash-abliterated

Stop any existing Splash server before starting another. Allow approximately 17.4 GB for the package download. These are Splash-specific packed binaries, not a Transformers, GGUF, or MLX checkpoint.

Once the server reports ready, open http://127.0.0.1:8000 or use its local API:

curl http://127.0.0.1:8000/v1/chat/completions \
  -H 'Content-Type: application/json' \
  -d '{
    "model": "audreyt/Qwen3.8-27B-Splash-abliterated",
    "messages": [{"role": "user", "content": "Suggest a short title for a community garden newsletter."}],
    "reasoning_effort": "none",
    "max_tokens": 128
  }'

Splash binds to loopback and disables authentication by default. Keep it local unless you have configured authentication and access controls. Require confirmation before giving any model access to files, accounts, or external actions.

Small evaluation, including failures

Checked on 2026-09-19 with Splash 1.0, an Apple M5 Max, 128 GB unified memory, and macOS 27.0. Requests used temperature 0 and reasoning_effort: none, except the rate problem, which used low. These are short smoke checks, not a benchmark or safety evaluation.

Counting tool selection and the tool-result response separately, 9 of 11 automated checks passed. Manual review also found a faithful Traditional Chinese translation and coherent two-sentence prose. A brief first-aid answer included cooling and escalation advice, but gave an imprecise cooling duration; this was not a clinical validation.

A limited stock-model comparison used the same two math prompts and settings. Stock Splash also returned 887 on the arithmetic prompt, but answered the rate problem correctly with 63. This one comparison is a warning about possible quality loss, not an estimate of its frequency. Use a calculator or verified code for numerical work.

The server advertised a 262,144-token context window on this machine. These checks did not test that capacity, long-context accuracy, general agent reliability, or speculative-decoding performance. They also do not establish that refusals have disappeared or that outputs are safe.

Construction and integrity

Starting package: incoai/Qwen3.8-27B-Splash, revision 9d27070b71f7142c6b6025f03ac011d70a73cb48.

For 70 output-projection tensors in layers 17–51, using zero-based file indices, the conversion adds the BF16 huihui-minus-Qwen weight difference to the dequantized Splash tensor, then requantizes to group-64 affine Q4. The changed sections are mlp-down plus either gdn-output or attention-output. This introduces additional quantization error; it is not numerically identical to huihui's BF16 model.

The corrected build preserves all other layer bytes and all other artifacts, including the draft, vision encoder, tokenizer, embedding, and head. The draft has not been retrained for the modified target. Upstream component references in the manifest describe the source package; this section describes the subsequent edits.

Verification covered SHA-256 and size for all 78 artifacts, byte preservation outside the intended sections, and six projection spot checks across GDN and full-attention layers. The six requantized sections matched the intended conversion byte-for-byte. Numerical consistency does not prove model quality.

Packaging correction: the initial upload, revision 1573dd2556bec088041334f40a44ab5e3d370aee, contained unintended zeroed bytes in 26 GDN input projections. This revision restores those bytes from the original package and updates the artifact hashes. The evaluation above applies to the corrected package. Use the current revision rather than that initial upload.

Use and limitations

Review generated material before relying on or sharing it. Weight edits can make the model more willing to answer without making the answer more accurate or appropriate. It may still refuse, invent facts, or provide poor advice. Avoid unsupervised high-stakes decisions and public deployment without an application-specific evaluation and safeguards.

Credits and license

Apache-2.0; see LICENSE. Credit goes to the Qwen team, mlx-community, huihui-ai, and Inco AI for the underlying model, conversion, weight edits, and Splash package and draft. Audrey Tang assembled this derivative package and made the packaging correction described above.