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.
| Model | huihui-ai/Huihui-Qwen3.8-27B-abliterated, dense 27B, hybrid Gated DeltaNet + attention |
| Target precision | 8-bit, GGUF Q8_0 (29.0 GB) |
| Speculative decoding | DFlash2 block-diffusion draft by Inco AI |
| Vision | Yes (Qwen3.8 vision tower, unchanged by abliteration) |
| Context | 262,144 tokens |
| Package size | 34.8 GB |
| Runtime | LM Studio with the Splash runtime (Apple silicon) |
| Licence | Apache 2.0 |
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).
| Profile | reasoning_effort |
Temperature | Top P | Top K | Best for |
|---|---|---|---|---|---|
| Fast (benchmarked) | low |
0.6 | 0.95 | 20 | Everyday chat, documents and RAG, tool calling, code. These are the settings behind every score on this card. |
| Balanced | medium |
1.0 | 0.95 | 20 | General use where you want more deliberate answers without the full thinking cost. |
| Maximum quality | xhigh (template default) |
1.0 | 0.95 | 20 | Hard maths, science and multi-step reasoning. Allow at least 32K output tokens. |
The 1.0 / 0.95 / 20 sampling values are Qwen's thinking-mode recommendation, embedded in the GGUF's own metadata.
| Setting | Value | Why |
|---|---|---|
| Min P | off (0) | Not used by Qwen's recipe |
| Repeat penalty | off | Penalising repeated tokens damages code, JSON and tool-call arguments |
| Presence penalty | off (0) | Degrades long, structured documents |
| Context length | 32K-98K | Tested at 98,304 tokens; the model supports 262,144 |
| Parallel requests | up to 4 | The Splash engine batches up to four sequences |
| Max response tokens | unlimited, or at least 16K | Thinking counts towards the response; maths problems used up to 32K |
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}
]}
}
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: ..."}]
}'
low used 27% fewer tokens
than xhigh on the same maths problem, and reached the same answer.preserve_thinking), so long multi-turn chats fill the context
faster. Start a new chat for unrelated tasks.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.
| Test | Questions | This package Huihui Q8_0 + DFlash2 (Splash) |
Huihui Q6_K_L GGUF (llama.cpp, MTP) |
Qwen3.8-27B Q8_0 + DFlash2 (Splash) |
Qwen3.8-27B MLX 8-bit |
|---|---|---|---|---|---|
| Tool calling | 20 | 20/20 | 20/20 | 20/20 | 20/20 |
| Document grounding | 40 | 38/40 | 39/40 | 39/40 | 39/40 |
| Instruction following | 50 | 46/50 | 47/50 | 47/50 | 45/50 |
| Code | 30 | 28/30 | 30/30 | 28/30 | 27/30 |
| Vision | 20 | 19/20 | 19/20 | 19/20 | 20/20 |
| Long context | 3 | 3/3 | 3/3 | 3/3 | 3/3 |
| Maths | 8 | 7/8 | 8/8 | 7/8 | 7/8 |
| Science | 25 | 22/25 | 21/25 | 22/25 | 23/25 |
| Overall (mean of the 7 text tests, out of 10) | 9.37 | 9.66 | 9.44 | 9.40 |
The tests:
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.
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.
| Throughput (tokens/s) | This package Huihui Q8_0 + DFlash2 (Splash) |
Huihui Q6_K_L GGUF (llama.cpp, MTP) |
Qwen3.8-27B Q8_0 + DFlash2 (Splash) |
Qwen3.8-27B MLX 8-bit |
|---|---|---|---|---|
| Writing (prose) | 36.1 | 33.5 | 40.5 | 18.2 |
| Code | 78.7 | 32.3 | 84.7 | 18.2 |
| Prompt processing | 852 | 626 | 890 | 805 |
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.
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:
| Test | This package Huihui Q8_0 + DFlash2 (Splash) |
Huihui Q6_K_L GGUF (llama.cpp, MTP) |
Qwen3.8-27B Q8_0 + DFlash2 (Splash) |
Qwen3.8-27B MLX 8-bit |
Median time per answer (this package) |
|---|---|---|---|---|---|
| Tool calling | 46.2 | 36.4 | 50.1 | 16.4 | 2 s |
| Document grounding | 75.3 | 34.2 | 76.8 | 17.4 | 7 s |
| Instruction following | 59.8 | 37.1 | 60.3 | 18.5 | 19 s |
| Code | 100.2 | 35.6 | 94.7 | 18.4 | 8 s |
| Vision | 74.9 | 43.5 | 73.5 | 17.8 | 4 s |
| Long context | 14.1 | 12 | 20.1 | 9.7 | 8 s |
| Maths | 73.8 | 31.7 | 71.7 | 17.9 | 65 s |
| Science | 79 | 33.3 | 75.1 | 18.1 | 31 s |
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.
| Part | Source | Notes |
|---|---|---|
| Target model | huihui-ai/Huihui-Qwen3.8-27B-abliterated-GGUF @ 1c69098, Huihui-Qwen3.8-27B-abliterated-Q8_0.gguf |
unchanged |
| Draft model | incoai/Qwen3.8-27B-DFlash2 @ 015e795 |
unchanged |
| Vision tower | mmproj-BF16.gguf from unsloth/Qwen3.8-27B-GGUF @ 4ca7207; vision/model.bin from incoai/Qwen3.8-27B-Splash |
unchanged; abliteration does not modify the vision tower |
| Tokenizer, chat template | incoai/Qwen3.8-27B-Splash | unchanged; the tokenizer is identical to the target GGUF's |
config.json, model.json, manifest.json |
adapted from dmeight/Qwen3.8-27B-Splash-8bit | target source, digests and artifact list updated |
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.
Two document-grounding behaviours repeated on every attempt, at reasoning effort low and xhigh alike:
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.
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.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.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 | Size | SHA-256 |
|---|---|---|
config.json |
651 B | 806669083aa894f1e68654ffb354cf5e43ef33d89fa4e6327cdecb526a3592f1 |
draft/config.json |
1,239 B | 873e3556509b0da06e29654ba00d4944888d4b5e8a33afde25f7eb27d321e980 |
draft/model.safetensors |
3.85 GB | 67fc76d68dc5a9415511a4f394ef744d67510cd20e93b37cc2cc7d28e4bab65c |
model.json |
539 B | 5983186f1b6b6256495e9a8a23764d55b8f7b2feea09d2b63b69803a9f82404d |
target/Huihui-Qwen3.8-27B-abliterated-Q8_0.gguf |
29.05 GB | 427b9416c2aba9d21da4e90ddac50a2bc8b6d0bfef675af3580a43aa38e000c3 |
tokenizer/chat_template.jinja |
8,940 B | 3b3b38b6ea9eb7df57eb47892dfd15be9f84361722b229b9220654a8cf467f74 |
tokenizer/config.json |
4,932 B | 14b65a0ee06517060a6bbd979bb1a8ff54e7b304b1a1f01d54344b88b8285e85 |
tokenizer/tokenizer.json |
19.99 MB | 06b9509352d2af50381ab2247e083b80d32d5c0aba91c272ca9ff729b6a0e523 |
tokenizer/tokenizer_config.json |
1,165 B | 792fa3f0cb88b111e54ef3134c873531008c4df471d108da17903426e308aa7b |
tokenizer/vocab.json |
6.72 MB | ce99b4cb2983d118806ce0a8b777a35b093e2000a503ebde25853284c9dfa003 |
vision/mmproj.gguf |
0.93 GB | 83ee4f4f205fa514161778c41df1ea14144faa0f713510893b63c2395f5c2d53 |
vision/model.bin |
0.93 GB | 8973858e75ec3464f626d5f9111dbceda58f3bac5737ea5693767686003c65df |
Packaged and benchmarked by SilverSurfer79.
Apache 2.0. Every component is licensed under Apache 2.0; see LICENSE and NOTICE.
huihui-ai/Huihui-Qwen3.8-27B-abliterated