license: apache-2.0
library_name: splash
pipeline_tag: image-text-to-text
base_model:

• huihui-ai/Huihui-Qwen3.8-27B-abliterated
  base_model_relation: quantized
  language:
• en
  tags:
• splash
• lm-studio
• apple-silicon
• metal
• local-inference
• dflash2
• speculative-decoding
• qwen3.8
• abliterated
• uncensored
• vision
• 8-bit
• gguf

Huihui-Qwen3.8-27B-abliterated · Splash 8-bit

huihui-ai's abliterated Qwen3.8-27B at 8-bit (GGUF Q8_0), packaged for the Splash runtime in LM Studio on Apple
silicon, with DFlash2 speculative decoding and vision.

At the time of writing, every other Splash build of an abliterated Qwen3.8-27B uses a 4-bit target. This package keeps huihui-ai's 8-bit
weights and adds the DFlash2 draft model, which proposes several tokens at a time for the 8-bit model to verify.
The output follows the 8-bit model (the draft changes speed, not the result). On the test Mac it answered the
benchmark's tasks 1.2-2.8x faster than huihui-ai's own Q6_K_L GGUF in
LM Studio (2.2x on document questions, 2.8x on code), scoring within two questions
of it on every test.

Unofficial and independent. Not made or endorsed by huihui-ai, Inco AI (Splash, DFlash2), LM Studio,
Alibaba Cloud (Qwen) or Unsloth. No weights were changed: every weight file is a byte-for-byte copy of its source.

At a glance

Requirements

• Apple silicon M3 or newer, macOS 26.4 or later (Splash's requirements).
• LM Studio with the Splash runtime installed and selected (Settings → Runtime). Tested with LM Studio
  0.4.28 and Splash runtime 1.3.0.
• About 35 GB of memory for the model at a 32K context; a 48 GB Mac or larger is recommended.

Install

From LM Studio: search for SilverSurfer79/Huihui-Qwen3.8-27B-abliterated-Splash-8bit in the model search and download it. It appears as
Huihui Qwen3.8 27B Abliterated Splash 8bit.

From the command line:

hf download SilverSurfer79/Huihui-Qwen3.8-27B-abliterated-Splash-8bit \
  --local-dir ~/.lmstudio/models/SilverSurfer79/Huihui-Qwen3.8-27B-abliterated-Splash-8bit

Then load it in LM Studio, or with lms load (the model key is derived from the folder name).

Recommended settings

Profiles

The 1.0 / 0.95 / 20 sampling values are Qwen's thinking-mode recommendation, embedded in the GGUF's own metadata.

Settings for every profile

LM Studio preset

Save this as ~/.lmstudio/config-presets/Huihui Qwen3.8 27B Splash.preset.json, then pick
Huihui Qwen3.8 27B Splash from the preset menu in the chat sidebar:

{
  "identifier": "@local:huihui-qwen3.8-27b-splash",
  "name": "Huihui Qwen3.8 27B Splash",
  "changed": true,
  "operation": {"fields": [
      {"key": "llm.prediction.temperature", "value": 1.0},
      {"key": "llm.prediction.topKSampling", "value": 20},
      {"key": "llm.prediction.topPSampling", "value": {"checked": true, "value": 0.95}},
      {"key": "llm.prediction.minPSampling", "value": {"checked": false, "value": 0}},
      {"key": "llm.prediction.repeatPenalty", "value": {"checked": false, "value": 1.1}},
      {"key": "llm.prediction.llama.presencePenalty", "value": {"checked": false, "value": 0}},
      {"key": "llm.prediction.maxPredictedTokens", "value": {"checked": false, "value": 32768}}
  ]},
  "load": {"fields": [
      {"key": "llm.load.contextLength", "value": 65536}
  ]}
}

Choosing the reasoning effort

The chat template reads reasoning_effort (low, medium or xhigh). LM Studio's chat window has no effort
selector for Splash packages, so chats there run at the default, xhigh. To choose per request, call LM Studio's
OpenAI-compatible server (Developer → Start Server):

curl http://localhost:1234/v1/chat/completions \
  -H 'Content-Type: application/json' \
  -d '{
    "model": "huihui-qwen3.8-27b-abliterated-splash",
    "reasoning_effort": "low",
    "temperature": 0.6, "top_p": 0.95, "top_k": 20,
    "messages": [{"role": "user", "content": "Summarise this policy in three paragraphs: ..."}]
  }'

Tips

• Reasoning effort is the main speed control. On the stock 8-bit Splash build, low used 27% fewer tokens
  than xhigh on the same maths problem, and reached the same answer.
• Describe length by structure, not word count. Asked for "about 300 words", Qwen3.8 counts words inside
  its thinking and can spend its budget doing so. Ask for "three paragraphs" or "a one-page summary" instead.
• Code and structured output are the fastest. The DFlash2 draft predicts them well, so they run at more than
  twice the speed of free prose.
• Thinking stays in the history by default (preserve_thinking), so long multi-turn chats fill the context
  faster. Start a new chat for unrelated tasks.
• System messages can appear anywhere in the conversation. The official Splash chat template accepts them
  after the first turn, which RAG and agent tools that inject context mid-conversation rely on.

Benchmarks

Measured locally on an Apple M5 Max with 128 GB, macOS 27.0.1, LM Studio 0.4.28, November 2026. Every model answered
the same questions with the same settings: reasoning effort low, temperature 0.6, top-p 0.95, top-k 20, 98,304-token
context, one request at a time, and no other model loaded. Every answer is graded exactly (executed code, official
IFEval checkers, numeric or letter answers); a missing or unparseable answer counts as wrong.

The tests:

• Tool calling (20): single, parallel and no-call function-calling cases.
• Document grounding (40): questions over supplied documents, including absent facts and false premises.
• Instruction following (50): IFEval prompts, Google's official checkers.
• Code (30): HumanEval+ functions, executed against the extended tests.
• Vision (20): tables, charts, an invoice and a degraded scan.
• Long context (3): retrieval and reasoning over ~30K-token logs.
• Maths (8): AIME 2025 problems, exact integer answers.
• Science (25): GPQA Diamond multiple choice.

As a consistency check, every question this package missed was asked again: with the same settings, 5 of 13 passed; at reasoning effort xhigh, 4 of 13 passed. The table shows first attempts at effort low only, as for every other model.

The question sets are small, so treat differences of one or two questions as noise. The comparison columns are
huihui-ai's Q6_K_L GGUF of the same model run by LM Studio's llama.cpp engine, and the stock (non-abliterated)
Qwen3.8-27B as an 8-bit Splash package and as LM Studio's MLX 8-bit build.

Speed on the test Mac

MacBook Pro, Apple M5 Max (18-core CPU, 40-core GPU), 128 GB unified memory, macOS 27.0.1, LM Studio 0.4.28,
Splash runtime 1.3.0. One request at a time, no other model loaded.

Fixed prompts

Writing and code were timed on fixed prompts; prompt processing on a ~16K-token prompt. Speculative decoding helps
most where text is predictable, so code gains far more than prose. The DFlash2 draft was trained against the stock
Qwen3.8-27B, so it agrees slightly less often with the abliterated target, which is why this package is a little
slower than the stock 8-bit Splash build.

Real workloads

Tokens per second across every benchmark answer, from sending the request to the last token (prompt processing and
thinking included), with the median time this package took per answer:

Memory

LM Studio reports 32.4 GiB for the loaded model at a 39,884-token context: the 8-bit target, the draft, the
vision tower and the KV cache. Loading reads the 35 GB package from disk, so the first load after a restart takes
longest.

What's inside

The target has the same architecture, hyperparameters and tokenizer as the stock Qwen3.8-27B target, so the
configuration and tokenizer files carry over unchanged. NOTICE lists every source and change.

Observed in testing

Two document-grounding behaviours repeated on every attempt, at reasoning effort low and xhigh alike:

• Filling a gap with a nearby figure. Asked for a rate the documents do not state, it quoted a related rate
  from the same documents. The stock Qwen3.8-27B MLX build gave the identical answer, so this comes from the base
  model, not from this package.
• Accepting a false premise. When a question asserted a wrong figure, it answered from the question's figure
  instead of correcting it from the documents. huihui-ai's Q6_K_L build did the same on a different false-premise
  question.

For document work, ask the model to quote the passage behind every figure it gives, and to say plainly when the
documents do not contain the answer. Both instructions make these errors easy to catch.

Known issues

• LM Studio labels it "4bit". LM Studio gives every Splash package a fixed 4-bit label. The target here is
  8-bit Q8_0, as the 29 GB target file shows.
• It relies on the current package layout. LM Studio lists a folder as a Splash model from its manifest.json,
  and the Splash engine loads the GGUF target through model.json. A future LM Studio or Splash runtime may change
  this; once LM Studio's Splash runtime loads 8-bit GGUF models directly, the original sources can be used instead.

Without LM Studio

• llama.cpp, Ollama and other GGUF tools can use target/Huihui-Qwen3.8-27B-abliterated-Q8_0.gguf and
  vision/mmproj.gguf as they are, without the DFlash2 speed-up. It is simpler to download them from
  huihui-ai/Huihui-Qwen3.8-27B-abliterated-GGUF.
• MLX tools cannot use it: the weights are GGUF.

Responsible use

This model inherits huihui-ai's abliteration, which removes much of the base model's refusal behaviour. It can
produce sensitive, controversial or inappropriate content, and it has not been through safety tuning. Review its
output, do not deploy it in public-facing products without your own safeguards, and make sure your use complies with
the law and with the licences of the upstream works. See huihui-ai's model card for their full usage warnings.

File integrity

Credits

• Qwen3.8-27B by the Qwen team, Alibaba Cloud.
• Abliteration and GGUF quantization by huihui-ai.
• Splash engine, package format, DFlash2 draft, tokenizer and vision files by Inco AI.
• Vision projector GGUF by Unsloth.
• 8-bit Splash package layout by dmeight.

Packaged and benchmarked by SilverSurfer79.

Licence

Apache 2.0. Every component is licensed under Apache 2.0; see LICENSE and NOTICE.