Qwen3.8 Flash-Next-Uncensored-MC

Qwen3.8-Flash-Next-Uncensored, OrcaRouter's abliterated BF16 build of Qwen's Qwen3.8-Flash-Next (177B parameters, 512 experts per layer, 10 active), for one NVIDIA RTX 5090 in the file format of MegaCapybara. Made the same way as perkel/Qwen3.8-Flash-Next-MC, in the same two models: Tiny (routed experts at 2 bits: faster, more context) and Small (routed experts at 3 bits: closer to the original). Every model file carries its measured accuracy and speed.

Safety alignment largely removed. The source model had its refusal behaviour removed by abliteration. It follows requests that the original Qwen3.8-Flash-Next would refuse, including harmful ones, and has no built-in guardrails.

It is meant for research, red-teaming, refusal and interpretability studies, and controlled experiments. Do not serve it to others without your own safety and moderation layers. You are responsible for how you use it and for what it generates, under the license below and the laws that apply to you.

Engine support: MegaCapybara 1.36 or later; its launcher downloads a model with its parts and the DFlash2 drafter from this repository.

How it was made

The abliteration removed one direction from the matrices that write to the residual stream (attention and DeltaNet outputs, the experts' and shared expert's down projections, the embedding); the routers, the experts' gate and up projections, the n-gram table, the output head and the vision tower are untouched. So only the edited matrices were quantized again, from OrcaRouter's BF16 weights with exllamav3's EXL3 converter and its calibration (the same tool and settings as turboderp's release of the original); everything else is the original release's, bit for bit:

  • Tiny: the routed experts' down projections at 2 bits, the attention / DeltaNet outputs and shared experts' down projections at 4 bits, quantized here; the other tensors from turboderp's EXL3 2.05 bpw.
  • Small: the routed experts at 3 bits from Lygodactylus's EXL3 3.05 bpw of the same uncensored model (layer 0's experts quantized here at 3 bits; that release has them at 4); the rest as Tiny.
  • MTP layer re-quantized at 4 bits from the uncensored BF16 release (its own outputs were edited too).

Files

File What it is On disk
qwen3.8-flash-next-uncensored-tiny.mcapy Qwen3.8 Flash-Next-Uncensored-MCTiny: a model file (its routed experts at 2 bits, scores, default expert profile) 30.7 GB
qwen3.8-flash-next-uncensored-small.mcapy Qwen3.8 Flash-Next-Uncensored-MCSmall: a model file (its routed experts at 3 bits, scores, default expert profile) 45.8 GB
qwen3.8-flash-next-uncensored-trunk.mcapy the dense layers, gated residual, routers, head and embedding (4 bits), tokenizer and config 5.1 GB
qwen3.8-flash-next-uncensored-mtp.mcapy the MTP layer at 4 bits (the speculative drafter) 1.4 GB
qwen3.8-flash-next-dflash2.mcapy DFlash2 speculative drafter, converted from incoai/Qwen3.8-Flash-Next-DFlash2 (the default when present: the same answers, faster than the MTP layer) 0.61 GB
qwen3.8-flash-next-ngram.mcapy the n-gram embedding table (4 bits): the same file as in Qwen3.8-Flash-Next-MC 26.3 GB
qwen3.8-flash-next-vision-fp8.mcapy the vision tower in FP8 (images): the same file as in Qwen3.8-Flash-Next-MC 0.56 GB
qwen3.8-flash-next-vision-bf16.mcapy the vision tower in BF16: the same file as in Qwen3.8-Flash-Next-MC 0.99 GB

Each model file names its parts; put them in the same folder. If you already have Qwen3.8-Flash-Next-MC, the n-gram and vision files are already there (MegaCapybara's launcher recognises them and skips them): Tiny then adds 37 GB, Small 52 GB. On RAM and VRAM the models behave as the originals (see their card): plan on 64 GB of RAM.

Accuracy against the uncensored original

KL divergence and top-1 agreement against OrcaRouter's BF16 weights computed with FP32 math, by MegaCapybara's engine, KV cache 8/8, no YaRN (in brackets: the original models against Qwen's BF16, for comparison):

Text Tiny KL Tiny top-1 Small KL Small top-1
Held-out set (81,880 tokens: reasoning traces, math, code, chat, wiki) 0.1464 (0.1454) 86.5% (86.7%) 0.0540 (0.0613) 91.7% (91.7%)
Chat and tool calls written by Flash-Next itself (the model's turns) 0.0752 (0.0760) 89.9% (90.0%) 0.0309 (0.0283) 93.6% (93.7%)
Real agentic coding sessions (the model's own new turns) 0.2247 (0.2234) 85.0% (85.2%) 0.1276 (0.1260) 88.6% (88.7%)

Per source of the held-out set, Tiny / Small: reasoning 0.094 / 0.040, math 0.059 / 0.021, code 0.093 / 0.041, chat 0.326 / 0.104, wiki 0.160 / 0.065.

At 32K tokens (math + code KL: Tiny 0.052, Small 0.020), against KV 8/8 without YaRN (relative KL; math + code / agentic coding):

Setting Tiny Small
KV 6/6 none measurable +0.8% / none measurable
KV 6/4 +1.4% / +3.8% +8.3% / +3.8%
KV 4/6 +2.6% / +0.9% +9.8% / none measurable
KV 4/4 +5.8% / +3.2% +14.1% / +3.2%
YaRN x2 (512K-token conversations) +4.3% / +3.9% +9.7% / +1.9%
YaRN x4 (1M-token conversations) +14.1% / +8.2% +30.7% / +10.8%

(Repeat runs of the same setting differ by up to ~3%.)

Speed (RTX 5090)

Through MegaCapybara's server, one conversation unless stated, greedy, thinking off, answers up to 2,000 tokens (tokens per second; MTP: the model's own MTP layer drafting, the answers unchanged):

Model Experts in VRAM MTP, code MTP, prose MTP, 4 / 8 conversations at once No speculation, code / prose Reading an 8K prompt
Tiny ~20,000 (small context) 466 325 849 / 1,050 296 / 288 8,300
Tiny 9,863 226 209 270 / 265 182 / 198 8,200
Tiny 4,220 130 125 150 / 162 109 / 120 7,500
Small ~13,000 (small context) 268 221 257 / 272 210 / 214 7,800
Small 10,513 212 189 196 / 204 164 / 174 7,200

The launcher projects each setting's speed from these points by how many experts it leaves in VRAM.

With the DFlash2 drafter (qwen3.8-flash-next-dflash2.mcapy, the default when it is in the folder), Tiny with ~20,000 experts in VRAM, measured in one session against the MTP layer: code 483 tok/s (MTP 418), prose 316 (302), 4 / 8 conversations at once 722 / 920 (714 / 1,018). Its gain is largest on the original model; on the uncensored one the drafter (trained on the original's outputs) keeps fewer drafts with several conversations at once.

License

Qwen Community License 1.0, as Qwen3.8-Flash-Next, from which every weight here derives (OrcaRouter's card states Apache 2.0; the original's license is the one kept here). Note its conditions: a commercial "Model as a Service" or "AI Work Assistant" business needs a separate license from Qwen; products past 100M monthly users or US$20M monthly revenue must show the model name. What was changed and by whom: NOTICE. The DFlash2 drafter is Apache 2.0, as incoai's release. The MegaCapybara engine is MIT licensed.

Downloads last month
103
Inference Providers NEW
This model isn't deployed by any Inference Provider. ๐Ÿ™‹ Ask for provider support

Model tree for perkel/Qwen3.8-Flash-Next-Uncensored-MC

Quantized
(55)
this model