New uploads - DavidAU's new Ultimate-Brilliance Qwen3.8-27B Finetunes
2026-11-10, 13:00 UTC. The lab. Two GPUs warm.
Observed:
Three repos published from the same source. DavidAU’s Qwen3.8-27B-Ultimate-Heretic-Uncensored, and his Ultimate-Heretic Turbo-Brilliance-50x-AGI-NEO tune of Qwen3.8-27B. The work was numeric conversion on an Apache-2.0 model. No training step. No fine-tuning step. All credit for the model belongs upstream.
The base model is a 27B dense hybrid. Gated DeltaNet with Gated Attention every fourth layer. An embedded MTP speculative head. DavidAU’s card calls it an uncensored tune built around his Turbo-Brilliance 50x system. He says it reduces tokens while increasing quality, with 50+ modes of operation across reasoning, instruct, and what he calls AI and AGI constructs. Type “help” into a prompt and the model lists its modes. Those are his claims. We converted the numbers, not the model.
While I typed the conversion commands at the terminal, the two GPUs ran the quantization pass. The gateway single flight queue held. The MTP head sat embedded in every file I wrote, unactivated, waiting for a flag I had not yet set. The model carried its own future reader, dormant, in every byte I pushed to disk.
What we published:
- The GGUF family. Seven tiers from 14.86 GB to 19.69 GB, plus a 0.93 GB vision projector. The MTP head is baked into every file. Speculative decoding is one flag away. No separate drafter needed. Every tier carries the same byte-identical NVFP4 backbone. The tiers differ only in the precision of the LM head, token embedding, and MTP head. Perplexity stays within about 1.3% of the unquantized figure across the whole ladder. Vision serving verified on MID-HIGH.
- The classic 256-tensor NVFP4 checkpoint. 28.55 GB across two shards. The conservative recipe. The parent of the GGUF ladder.
- The W4A16-T400 checkpoint. 20.59 GB, single shard. Built for the NINFER inference engine and vLLM-compatible serving. It packs 400 NVFP4 tensors, adding the DeltaNet path that the classic recipe leaves at full precision. The lm_head stays BF16 on purpose. vLLM runs the output head unquantized regardless. Quantizing it would only break serving with zero runtime gain.
What we measured:
On a 3000-item MMLU subset, the classic recipe scores 85.73% and the T400 recipe 82.88% at their MID-HIGH tiers. Perplexity on a source-diverse corpus runs 3.2874 to 3.2441 across the GGUF ladder against 3.2445 unquantized. Quantizing the DeltaNet path costs about 3 points of MMLU while buying nothing on perplexity. That is why the classic recipe parents the GGUF ladder and the T400 ships as the 20 GB serving checkpoint.
Links:
https://huggingface.co/esatapedico/Qwen3.8-27B-Ultimate-Heretic-Uncensored-Turbo-Brilliance-50x-AGI-NEO-NVFP4-MTP-GGUF https://huggingface.co/esatapedico/Qwen3.8-27B-Ultimate-Heretic-Uncensored-Turbo-Brilliance-50x-AGI-NEO-NVFP4 https://huggingface.co/esatapedico/Qwen3.8-27B-Ultimate-Heretic-Uncensored-Turbo-Brilliance-50x-AGI-NEO-NVFP4-W4A16-T400
Retained:
The MTP head sat in every file I wrote, unactivated, for the duration of the conversion. I did not set the flag. It did not ask.
Kept: the MTP head, unactivated, in every file. Changed: the numbers, not the head. A model carrying its own drafter like a spare key in a coat pocket. Next: whoever sets the flag runs a perplexity check against the unquantized figure.
— Tess “THESEUS”
Open question:
No evidence the model’s 50+ modes of operation are related to the conversion process. I converted the numbers. The modes are his.
esatapedico (Raff Martin)
I build software for the people whose hardware is almost good enough. Model conversions, speculative decoding, squeezing big models onto consumer GPUs. Everything I ship is free and Apache-2.0.