license: apache-2.0
base_model: DavidAU/Qwen3.8-27B-Ultimate-Heretic-Uncensored
pipeline_tag: text-generation
library_name: transformers
description: "W4A16 NVFP4 safetensors checkpoint of DavidAU's Ultimate-Heretic Turbo-Brilliance-50x-AGI-NEO tune of Qwen3.8-27B: 400 packed NVFP4 projections (T400) with lm_head, MTP head, vision tower, and DeltaNet dynamics in BF16. MTP in every name. Blackwell sm_120."
tags:
- nvfp4
- qwen3.8
- blackwell
- mtp
- speculative-decoding
- ultimate-heretic
- compressed-tensors
- vllm
- llama.cpp
language: - en
- multilingual
quantization_config:
quant_method: compressed-tensors
format: nvfp4-pack-quantized
Qwen3.8-27B-Ultimate-Heretic-Uncensored-Turbo-Brilliance-50x-AGI-NEO-NVFP4-W4A16-T400
NVFP4 checkpoint of DavidAU/Qwen3.8-27B-Ultimate-Heretic-Uncensored, DavidAU's Ultimate-Heretic Turbo-Brilliance-50x-AGI-NEO tune of Qwen3.8-27B, quantized for vLLM serving.
HF repo: esatapedico/Qwen3.8-27B-Ultimate-Heretic-Uncensored-Turbo-Brilliance-50x-AGI-NEO-NVFP4-W4A16-T400
NVFP4 here means W4A16 with FP8 E4M3 scales, group size 16, weight only, datafree (no calibration corpus). The W4A16-T400 suffix reads as: W4A16 (4-bit weights, FP8 E4M3 scales, group size 16, weight only) with T400 (400 packed NVFP4 tensors). The quantized set covers the MLP on all 64 layers, Q, K, V, O on the 16 full-attention layers, and the DeltaNet in_proj_qkv/z/out_proj tensors on the 48 linear-attention layers. The language model Linear projections are quantized; the vision tower, the MTP head, lm_head, the Gated DeltaNet conv1d and in_proj_a/in_proj_b entries, norms, and scales stay at full precision.
A companion checkpoint with the conservative 256-tensor recipe lives at esatapedico/Qwen3.8-27B-Ultimate-Heretic-Uncensored-Turbo-Brilliance-50x-AGI-NEO-NVFP4 with its own card.
Follow along & support
I post updates on new conversions, benchmarks, and what I'm working on over on my site. Follow along there to keep up with new releases and the work in progress. If you'd like to support more of it, you will find the options there. I do this on consumer hardware and like seeing how far it goes. More is on the way.
🌐 esatapedico.com. Updates, work-in-progress, and ways to support.
At a glance
| Field | Value |
|---|---|
| Format | compressed-tensors nvfp4-pack-quantized |
| Quantization | W4A16 ("NVFP4A16"), group size 16, FP8 E4M3 scales, weight only, 400 NVFP4 tensors |
| Calibration | none (datafree) |
| Quantized | MLP on all 64 layers + Q/K/V/O on the 16 full-attention layers + DeltaNet in_proj_qkv/z/out_proj on the 48 linear-attention layers (400 packed NVFP4 tensors) |
| Kept in BF16 | vision tower, MTP head (15 tensors), lm_head, Gated DeltaNet conv1d and in_proj_a/in_proj_b entries, norms, scales |
| Base | DavidAU/Qwen3.8-27B-Ultimate-Heretic-Uncensored (27B dense hybrid, Apache-2.0) |
Quantized set covers the MLP on all 64 layers (192 tensors) plus Q, K, V, O on the 16 full-attention layers (64 tensors) plus the DeltaNet in_proj_qkv/z/out_proj tensors on the 48 linear-attention layers (144 tensors). The lm_head stays at full precision on purpose: vLLM runs the output head unquantized regardless, and no vLLM release loads a compressed-tensors NVFP4 head, so quantizing it only breaks serving with zero runtime gain. Vision tower, MTP head, and the DeltaNet conv1d plus in_proj_a/b dynamics stay at full precision following the JC1DA exception pattern.
Serving
Built for vLLM and engines consuming vLLM-compatible compressed-tensors checkpoints, following the JC1DA-style exception pattern. Serve with --quantization compressed-tensors semantics (auto-detected from the config); FP8 KV remains a runtime flag, not checkpoint bytes. Keep vision intact and MTP for speculative decoding; no GGUF inside this repo.
Checkpoint
A single weight shard plus config and tokenizer files:
| File | Size | SHA-256 | Notes |
|---|---|---|---|
model.safetensors |
20,593,107,888 bytes (20.59 GB) | ba61939086e8d23df8acd5f9b8f5cfbb7fb5df20f39b2ddc4011d1bd937b3afb |
single weight shard, packed NVFP4 plus BF16 remainder |
model.safetensors.index.json |
171,032 bytes | b1f0ef6feb40acf7c604816d80901b986d05606f5188d81774a7d594c3eeb512 |
weight map, total size 20,593,107,840 |
config.json |
15,516 bytes | af7c0f83bdfe41f59ea70db8363467c24713bfc3187c33a12543b36f600a9705 |
architecture plus quantization_config |
tokenizer.json |
19,989,506 bytes (20.0 MB) | 2f4dd754486e96054d0c841ddde816748be0116cbbc8ba8b5311ae4853432a21 |
tokenizer |
tokenizer_config.json |
1,569 bytes | 13e78188447d1a36c7bc332a2c20f6eadb9823c255c5565c8107cf3e055111ca |
tokenizer config |
chat_template.jinja |
248,063 bytes | 73c64a8ea9af4491bda8b4f071964cf55dca998e8dbecea1538f1a0b8341dcf0 |
chat template |
generation_config.json |
219 bytes | 2c946d7c1ecbdfcb279a879795b7c28323119ade9b90b19f3170e0e1da2da8fa |
generation defaults |
recipe.yaml |
347 bytes | cc7e437c50a9af118ff9e5e35e58c3405506b78f30605a7a8bd98b80965a4f54 |
quantization recipe used |
Measurements
- Packed census: 400 of 400
Lineartargets quantized and packed, verified intact at save time. - MTP check: all 15 MTP tensors present in BF16, bit-identical to the source (14 versus the source shard plus
mtp.fc.weightversus the source MTP restoration file). - Preflight check: PASS.
- Total weight size: 20,593,107,888 bytes (20.59 GB) in one shard.
Usage
This checkpoint serves vLLM directly and is the quantization source for GGUF builds. The GGUF path is a converted file:
convert_hf_to_gguf.py --outtype auto
which keeps the native NVFP4 tensors intact. Requires a recent llama.cpp with NVFP4 (GGML type 40) support.
How this was made
- Quantized with a datafree NVFP4A16 recipe: targets
Linear, ignore listlm_head, visual tower, MTP head,conv1d,in_proj_a,in_proj_b. Sequential run, CPU only, no calibration corpus. - Packed 400 NVFP4 tensors. The 15 MTP tensors were restored to BF16 after quantization and verified bit-identical.
- Saved as a single safetensors shard with config, tokenizer, and recipe files.
Attribution & provenance
This is a derivative work built entirely from existing Apache-2.0 artifacts. Nothing here was trained or fine-tuned. This work covers the numeric conversion only and makes no ownership claims over the model. Credit belongs to:
- Alibaba / Qwen team for the base model family (Apache-2.0): the 27B dense hybrid architecture this tune builds on.
- DavidAU for the tune itself, Qwen3.8-27B-Ultimate-Heretic-Uncensored (Apache-2.0), including the Turbo-Brilliance-50x-AGI-NEO system this checkpoint is named for.
- This repo's author for the NVFP4 conversion and verification only.
License
Apache-2.0, identical to every upstream artifact. The base model license governs. Quantization is a transformation, not new training. When redistributing, please retain attribution to Qwen (Alibaba) and DavidAU as above.
"Qwen" is a trademark of Alibaba. Trademarks are used here only to identify upstream models. This repository is not affiliated with, sponsored by, or endorsed by Alibaba or DavidAU.
Card written with AI assistance.