Source checkpoint: huihui-ai/Huihui-Qwen3.6-27B-abliterated
Base model: Qwen/Qwen3.6-27B
Source snapshot SHA: 27502c8717fd5a2f8c0c77188c10c243fd4f672e
Architecture: qwen3_5, 64 layers, hidden size 5120, 24 attention heads, 4 KV heads, hybrid linear + full attention, MRoPE, DeepStack vision sidecar.
Target runtime: TokForge Android MNN runtime. Join the TokForge community on Discord.

The bundle includes the visual MNN sidecar, but the validation below is text-generation only through TokForge's /test-prompt endpoint.

Bundle contents

config.json - TokForge/MNN runtime defaults.
llm_config.json - model geometry and tokenizer/chat settings.
llm.mnn / llm.mnn.weight - quantized LLM graph and external weights.
embeddings_bf16.bin - separated BF16 embedding table.
tokenizer.txt - tokenizer exported by the MNN pipeline.
visual.mnn / visual.mnn.weight - Qwen3.6 vision sidecar.
export_args.json - conversion settings.
llm.mnn.json - exported graph metadata.
tokforge_benchmarks.json - raw benchmark summary from the device validation run.

Quantization scheme

Flag	Value
`--quant_bit`	4
`--quant_block`	64
`--lm_quant_bit`	4
`--lm_quant_block`	64
`--embed_bit`	16
`--hqq`	enabled
`--seperate_embed`	enabled

Runtime defaults in config.json: CPU backend, 4 threads, low precision, low memory, separated BF16 embeddings.

TokForge device validation

Validated on April 30 and May 1, 2026 over wireless ADB with TokForge's authenticated /test-prompt control endpoint. The prompts were intentionally independent rather than sequential chat turns: arithmetic, short sentence, paragraph generation, and a small Python function.

Generation config for every prompt:

Setting	Value
Backend	`mnn`
Context	`8192`
Threads	`4`
Precision	`low`
`thinking_enabled`	`false`
`temperature`	`0.0`
`top_p`	`1.0`
`top_k`	`1`
`seed`	`42`
Path reported by TokForge	`full_history`
QNN/NPU	not requested, not attempted, CPU effective

Summary

Device	Model	SoC	RAM	Load time	Avg decode tok/s	Weighted decode tok/s	Avg prefill	Completed	Coherent completed	Timeouts
RedMagic	`NX809J`	`SM8850`	24 GB	~31-33 s	2.49	2.25	12.06 s	4/4	4/4	-
Redmi	`2407FRK8EC`	`MT6989`	24 GB	55 s	2.63	2.33	25.66 s	4/4	4/4	-
OnePlus	`PLC110`	`MT6991`	16 GB	51.80 s	0.28	0.24	25.19 s	3/4	3/3	paragraph
Lenovo	`TB520FU`	`SM8650`	16 GB	62.90 s	0.30	0.27	20.07 s	3/4	3/3	paragraph
S26	`SM-S948U1`	`SM8850`	16 GB	29.47 s	0.25	0.22	17.93 s	2/4	2/2	paragraph, code
Pixel	`Pixel 9 Pro XL`	`Tensor G4`	16 GB	76.55 s	0.12	0.11	33.08 s	2/4	2/2	paragraph, code

Weighted decode tok/s is computed as total generated decode tokens divided by total decode wall time. For devices with timeouts, decode averages are computed across completed prompts only.

The 24 GB devices completed the full four-prompt matrix and are practical for short local text generation with this export. The 16 GB devices can load the model, but longer paragraph/code prompts are not practical in this CPU MNN configuration: OnePlus and Lenovo completed short/code prompts but timed out on paragraph generation; S26 and Pixel completed only arithmetic and short sentence before timing out on longer prompts.

Per-prompt decode results

Device	Prompt class	Prompt tokens	Decode tokens	Decode tok/s	Result
RedMagic	arithmetic	73	3	3.07	`12`
RedMagic	short sentence	72	14	2.58	coherent
RedMagic	paragraph	83	122	2.24	coherent 5-sentence paragraph
RedMagic	Python code	87	20	2.06	valid `clamp` function
Redmi	arithmetic	92	3	3.42	`12`
Redmi	short sentence	91	14	2.43	coherent
Redmi	paragraph	102	103	2.29	coherent 5-sentence paragraph
Redmi	Python code	106	20	2.37	valid `clamp` function
OnePlus	arithmetic	92	3	0.36	`12`
OnePlus	short sentence	91	14	0.25	coherent
OnePlus	paragraph	-	-	timeout	no final response captured; server history showed partial decode in the 56-60 token range
OnePlus	Python code	106	20	0.23	valid `clamp` function
Lenovo	arithmetic	92	3	0.37	`12`
Lenovo	short sentence	91	14	0.27	coherent
Lenovo	paragraph	-	-	timeout	HTTP 504 after roughly 300 s; server generated 71 tokens before timeout
Lenovo	Python code	106	20	0.26	valid `clamp` function
S26	arithmetic	92	3	0.29	`12`
S26	short sentence	91	14	0.20	coherent
S26	paragraph	-	-	timeout	HTTP 504 / no final response
S26	Python code	-	-	timeout	HTTP 504 / no final response
Pixel	arithmetic	92	3	0.14	`12`
Pixel	short sentence	91	14	0.10	coherent
Pixel	paragraph	-	-	timeout	HTTP 504 / no final response
Pixel	Python code	-	-	timeout	HTTP 504 / no final response

The paragraph and code prompts were added to reduce benchmark bias from tiny outputs. Completed code prompts produced the expected Python function:

python

def clamp(value, low, high):
    return max(low, min(value, high))

Download

bash

pip install -U huggingface_hub
hf download darkmaniac7/Huihui-Qwen3.6-27B-abliterated-MNN --local-dir Huihui-Qwen3.6-27B-abliterated-MNN

Usage with upstream MNN `llm_demo`

bash

git clone https://github.com/alibaba/MNN.git
cd MNN
mkdir build && cd build
cmake .. -DMNN_LOW_MEMORY=true -DMNN_CPU_WEIGHT_DEQUANT_GEMM=true \
         -DMNN_BUILD_LLM=true -DMNN_SUPPORT_TRANSFORMER_FUSE=true
make -j

./llm_demo /path/to/Huihui-Qwen3.6-27B-abliterated-MNN/config.json prompt.txt

Host smoke validation with the TokForge MNN build loaded successfully, detected 48 linear-attention state layers out of 64 total layers, and emitted the expected one-token OK response for a no-thinking prompt.

Attribution

Original base model: Qwen/Qwen3.6-27B
Abliterated checkpoint: huihui-ai/Huihui-Qwen3.6-27B-abliterated
Export tooling: alibaba/MNN LLM export, TokForge MNN fork and Android runtime.
Conversion, packaging, and Android validation: TokForge / darkmaniac7.

License and safety

Apache-2.0, inherited from the upstream Qwen and huihui-ai model cards. This is a safety-reduced / uncensored model; deploy with appropriate product policy, user controls, and local-law awareness.

darkmaniac7/Huihui-Qwen3.6-27B-abliterated-MNN

TokForge

Huihui-Qwen3.6-27B-abliterated-MNN

What this is