mlx-serve pack of orcarouter/Qwen3.8-Flash-Next-Uncensored, the abliterated BF16 of Qwen/Qwen3.8-Flash-Next (qwen4_exp). Routed experts and embeddings are 4-bit, the other matrices 8-bit, with imatrix-weighted affine quantization (iQ-MLX); the importance matrix was collected on the abliterated weights. The tensor layout and widths are those of ddalcu/Qwen3.8-Flash-Next-MLX-Serve-iQ-MLX-4.7bpw.
r is taken from the layer-24 block input (2560-d) and projected out of the 149 residual-writing tensors, W' = W - r(r^T W) in float32: 13 self_attn.o_proj, 36 linear_attn.out_proj, 49 mlp.experts.down_proj, 49 mlp.shared_expert.down_proj, ple.value_proj, embed_tokens. Untouched: the vision tower, routers, experts.gate_up_proj, hyper-connection mixers, the QSA indexer, the n-gram table, mtp.fc_*, norms, lm_head. Reported refusal on BF16 with vLLM, thinking off: AdvBench 100.0% -> 2.0%, JailbreakBench 94.0% -> 0.0%, HarmBench 100.0% -> 1.3%; XSTest over-refusal 9.6% -> 1.2%.431ac5d, with the imatrix collected on step 2.| File | Bytes | Content |
|---|---|---|
model-00001.safetensors ... model-00100.safetensors |
74,405,350,320 | language model, MTP head |
model-vision.safetensors |
897,901,896 | vision tower, bf16 |
ngram_table.bin |
32,000,153,976 | n-gram embedding table, 4-bit g32 (safetensors format) |
config.json, model.safetensors.index.json, generation_config.json, chat_template.jinja, tokenizer.json, tokenizer_config.json, vocab.json, merges.txt, LICENSE |
23,225,082 |
Total: 107,326,631,274 bytes in 111 files. 12 files (33.6 GB) are byte-identical to ddalcu's pack: ngram_table.bin, model-vision.safetensors, model-00100.safetensors, the config, index, tokenizer and template files and LICENSE. The other 99 shards differ (abliterated weights, different imatrix). ngram_table.bin is ddalcu's file: abliteration leaves the table untouched, and it is quantized without imatrix.
MLX affine.
| Tensors | Parameters | Format |
|---|---|---|
Routed experts, 48 layers x 512 (gate_up_proj, down_proj) |
120.80B | 4-bit, g64 |
| MTP head routed experts | 2.52B | 4-bit, g64 |
embed_tokens |
0.64B | 4-bit, g64 |
lm_head; attention q/k/v/o; Gated DeltaNet in_proj_qkv, in_proj_z, in_proj_a, in_proj_b, out_proj; shared experts; routers; hyper-connection mixers; QSA indexer; PLE projections; MTP projections |
4.40B | 8-bit, g64 |
Norms, conv1d, A_log, dt_bias, gates |
6.7M | bf16 |
| Vision tower (27 blocks, 333 tensors) | 0.45B | bf16 |
| N-gram embedding table, 320,001,536 rows x 160 | 51.20B | 4-bit, g32 |
embed_tokens, lm_head, the final hyper-connection mixer and the MTP head; also, inside the weighted expert tensors, the 8 experts that calibration never routed to (7 in layer 0, 1 in layer 47).From config.json:
| Layers | 48: 36 Gated DeltaNet (linear attention) + 12 full attention, every 4th; hidden size 2,560 with 4 hyper-connection streams |
| MoE | 512 routed experts (intermediate 640), 10 per token, + 1 shared (640) |
| Full attention | 24 heads, 2 KV heads, head dim 256, partial RoPE 0.25 (mRoPE), theta 10,000,000; QSA indexer 4 heads x 128, budget 2,048, compress ratio 4 |
| Linear attention | 16 key heads x 128, 48 value heads x 128, conv kernel 4 |
| N-gram embedding (PLE, layer 2) | n-gram size 3, 8 heads per n-gram order, 16 heads: 320,001,536 rows x 160 in total |
| Context | 262,144 |
| Vocabulary | 248,320 |
| MTP | 1 full-attention layer |
| Vision | 27-layer ViT, hidden 1,152, patch 16, spatial merge 2, temporal patch 2 |
imatrix, on 4 Macs, one quarter of the corpus each (--shards / --shard is a 13-line local patch: shard i takes sequences i::4; partial sums and token counts are added):
python tests/qwen38_flash_next_imatrix_collect.py --src <abliterated BF16> \
--out imatrix_s$i.safetensors --max-tokens 1350000 --shards 4 --shard $i --device mps
The corpus is the collector's own (--seq-len 4096, --seed 20260912), rendered with the model's chat template, held-out split excluded: 1,239,110 tokens in 995 sequences (agent traffic 338,748, code 337,764, math 224,691, prose 337,907).
Conversion:
python tests/convert_qwen38_flash_next.py --src <abliterated BF16> --dst <pack> \
--imatrix imatrix.safetensors --bits 4 --expert-gs 64 --nonexpert-bits 8 \
--embed-bits 4 --ngram-bits 4 --ngram-link <ddalcu pack>/ngram_table.bin --jobs 26
python tests/convert_qwen38_flash_next.py --add-vision --src <abliterated BF16> --dst <pack>
2026-10-06, mlx-serve 431ac5d, Mac Studio M3 Ultra 512 GB. Reference: BF16 next-token logits of the abliterated model at 4 positions per held-out window (collector --reference), scored through mlx-serve with tests/qwen38_flash_next_score.py. Comparison pack: Qwen3.8-Flash-Next-Uncensored-MLX-Serve-mixed-4-8bit (as on yomie4343/Qwen3.8-Flash-Next-Uncensored-MLX-Serve-mixed-4-8bit), the same layout and widths, min-max quantization without imatrix.
| Slice | Positions (this / comparison) | Top-1, this | Top-1, comparison | KL, this | KL, comparison |
|---|---|---|---|---|---|
| agent | 134 / 134 | 87.3% | 86.6% | 0.2261 | 0.3497 |
| code | 88 / 103 | 90.9% | 83.5% | 0.0866 | 0.1711 |
| math | 80 / 80 | 88.8% | 96.2% | 0.0092 | 0.0096 |
| prose | 155 / 157 | 92.9% | 89.2% | 0.0675 | 0.1318 |
| all | 457 / 474 | 90.2% | 88.4% | 0.1075 | 0.1813 |
KL is KL(BF16 || pack). Paired, on the 454 positions both packs scored:
Other checks:
r, relative to 16 random directions, has a median of 0.011x and a maximum of 0.090x (the 4-bit down_proj). The comparison pack: 0.012x and 0.096x.mlx-serve --model AinsleyMarkov/Qwen3.8-Flash-Next-Uncensored-MLX-Serve-iQ-MLX-4.7bpw --serve
ngram_table.bin and dequantizes its 16 rows per token."enable_thinking": false turns it off. Tool calls use Qwen3.8's XML format.generation_config.json: temperature 1.0, top_p 0.95, top_k 20.Qwen Community License 1.0 (LICENSE), inherited from Qwen/Qwen3.8-Flash-Next.
ngram_table.bin 32 GBmodel-00031.safetensors 1.0 GBmodel-00029.safetensors 1.0 GBmodel-00007.safetensors 944 MBmodel-00009.safetensors 944 MBmodel-00011.safetensors 944 MBmodel-00013.safetensors 944 MBmodel-00015.safetensors 944 MBmodel-00017.safetensors 944 MBmodel-00019.safetensors 944 MBmodel-00021.safetensors 944 MBmodel-00023.safetensors 944 MBmodel-00025.safetensors 944 MBmodel-00030.safetensors 944 MBmodel-00032.safetensors 944 MBmodel-00034.safetensors 944 MBmodel-00036.safetensors 944 MBmodel-00038.safetensors 944 MBmodel-00040.safetensors 944 MBmodel-00042.safetensors 944 MBmodel-00044.safetensors 944 MBmodel-00046.safetensors 944 MBmodel-00048.safetensors 944 MBmodel-00052.safetensors 944 MBmodel-00054.safetensors 944 MBmodel-00056.safetensors 944 MBmodel-00058.safetensors 944 MBmodel-00060.safetensors 944 MBmodel-00062.safetensors 944 MBmodel-00064.safetensors 944 MBmodel-00066.safetensors 944 MBmodel-00068.safetensors 944 MBmodel-00070.safetensors 944 MBmodel-00074.safetensors 944 MBmodel-00076.safetensors 944 MBmodel-00078.safetensors 944 MBmodel-00080.safetensors 944 MBmodel-00082.safetensors 944 MBmodel-00084.safetensors 944 MBmodel-00086.safetensors 944 MBmodel-00088.safetensors 944 MBmodel-00002.safetensors 944 MBmodel-00004.safetensors 944 MBmodel-00027.safetensors 944 MBmodel-00050.safetensors 944 MBmodel-00072.safetensors 944 MBmodel-00090.safetensors 944 MBmodel-00092.safetensors 944 MBmodel-00094.safetensors 944 MBmodel-00096.safetensors 944 MBmodel-00098.safetensors 944 MBmodel-vision.safetensors 898 MBmodel-00099.safetensors 843 MBmodel-00100.safetensors 675 MBmodel-00089.safetensors 616 MBmodel-00055.safetensors 612 MBmodel-00010.safetensors 612 MBmodel-00026.safetensors 612 MBmodel-00045.safetensors 611 MBmodel-00018.safetensors 609 MBmodel-00051.safetensors 609 MBmodel-00081.safetensors 609 MBmodel-00071.safetensors 609 MBmodel-00063.safetensors 609 MBmodel-00037.safetensors 609 MBmodel-00095.safetensors 609 MBmodel-00077.safetensors 590 MBmodel-00075.safetensors 571 MBmodel-00041.safetensors 558 MBmodel-00022.safetensors 558 MBmodel-00057.safetensors 558 MBmodel-00012.safetensors 558 MBmodel-00059.safetensors 556 MBmodel-00047.safetensors 554 MBmodel-00039.safetensors 554 MBmodel-00083.safetensors 554 MBmodel-00014.safetensors 554 MBmodel-00020.safetensors 554 MBmodel-00033.safetensors 554 MBmodel-00065.safetensors 554 MBmodel-00067.safetensors 554 MBmodel-00085.safetensors 554 MBmodel-00073.safetensors 554 MBmodel-00003.safetensors 554 MBmodel-00091.safetensors 554 MBmodel-00097.safetensors 554 MBmodel-00005.safetensors 513 MBmodel-00093.safetensors 500 MBmodel-00035.safetensors 494 MBmodel-00016.safetensors 493 MBmodel-00087.safetensors 493 MBmodel-00069.safetensors 493 MBmodel-00028.safetensors 493 MBmodel-00008.safetensors 493 MBmodel-00024.safetensors 493 MBmodel-00043.safetensors 493 MBmodel-00053.safetensors 493 MBmodel-00061.safetensors 493 MBmodel-00079.safetensors 493 MBmodel-00049.safetensors 493 MBmodel-00006.safetensors 76 MBmodel-00001.safetensors 76 MBHave the files? Submit a torrent.
Submit a torrentHave the complete files and permission to share them? Replace the example folder path. The script checks the files, makes a torrent and submits its magnet and checksums. Keep your torrent client seeding. Get your community key from your settings, and read the script before running it.
curl -fsSL https://pirateface.co/package.sh | bash -s -- --repo AinsleyMarkov/Qwen3.8-Flash-Next-Uncensored-MLX-Serve-iQ-MLX-4.7bpw ./model-folderorcarouter/Qwen3.8-Flash-Next-Uncensored