trithemius/Qwen3.6-27B-Fable-Fusion-711-Uncensored-Heretic-MTP-PrismAura-5.5bit

Qwen3.6-27B-Fable-Fusion-711-Uncensored-Heretic — PrismAura 5.5bit (Blackwell, NVFP4 + FP8 + BF16, MTP)

PrismaQuant export of DavidAU/Qwen3.6-27B-Fable-Fusion-711-Uncensored-Heretic-NM-DAU-MTP for vLLM compressed-tensors serving on NVIDIA Blackwell.

This is a PrismAura artifact: a per-Linear mixed-precision allocation chosen by AURA — PrismaQuant's production-faithful KL–Fisher allocator — over the full NVFP4 / FP8 / BF16 format menu. Rather than quantizing every Linear to a single fixed format, AURA prices each (Linear, format) pair by its measured second-order effect on the model's output distribution and solves a bit-budget knapsack over the whole network, spending precision where the model is sensitive and saving it where it isn't. MTP (multi-token prediction) tensors are included and left in BF16, so the model can be served with vLLM's speculative decoding out of the box.

The recipe mirrors rdtand/Qwen3.6-27B-PrismaAURA-5.5bit-vllm (same allocator, same target bit budget, same format menu), applied here to DavidAU's "Fable-Fusion-711" multi-stage fine-tune/merge of Qwen3.6-27B instead of the stock instruct checkpoint.

What this model is

Qwen3.6-27B-Fable-Fusion-711-Uncensored-Heretic-NM-DAU-MTP is DavidAU's multi-stage fine-tune, multi-fine-tune, and multi-stage merge built on top of Qwen/Qwen3.6-27B, combining contributions from DavidAU, Nightmedia, TeichAI (Polaris dataset), armand0e (Fable traces), and trohrbaugh. It was additionally decensored using Heretic v1.2.0+custom with the Arbitrary-Rank Ablation (ARA) method, which removes refusal behavior directly from the weights rather than through a system prompt or fine-tune. Per the base repo's own measurements, this drops refusals from 99/100 to 4/100 (KL divergence vs. the original Qwen3.6-27B: 0.0469).

This PrismAura build changes none of that behavior — it is a numerical re-encoding of the same weights into a mixed-precision compressed-tensors layout for efficient vLLM serving. It does not add, remove, or retrain anything at the weight level beyond quantization.

⚠️ This is a decensored, general-purpose model whose upstream tuning intentionally reduces refusal behavior, including for explicit, violent, and NSFW content. It carries no built-in content moderation. Deploy it behind your own usage policy and access controls if that matters for your use case.

Quantization details

• Source model: DavidAU/Qwen3.6-27B-Fable-Fusion-711-Uncensored-Heretic-NM-DAU-MTP (source weights identical to nightmedia/Qwen3.6-27B-Architect-Polaris2-Fable-B-F451)
• Base architecture: Qwen/Qwen3.6-27B — 27B dense, causal LM with vision encoder, 64 hidden layers, 262,144-token native context
• Method: PrismaQuant / AURA (KL–Fisher-priced, per-Linear mixed-precision allocation; multiple-choice knapsack over a {NVFP4, FP8_E4M3, BF16} format menu)
• Local rounding: deliberate per-Linear render under every format (GPTQ-style with damping, activation-order handling, scale-faithful NVFP4 export)
• Export format: vLLM compressed-tensors, mixed precision
• Target hardware: NVIDIA Blackwell (NVFP4-native execution)
• MTP tensors: included, BF16 passthrough
• Vision tower: included, BF16 passthrough (uniformly assigned, not run through the KL–Fisher allocator)
• Passthrough dtype policy: source dtype preserved (no silent FP32 upcasting)
• Target bit budget: 5.5 bits per quantizable parameter

Allocator search (bit budget sweep)

AURA was run across a sweep of target bit budgets against the source model, tracking the achieved bits/parameter and the resulting KL–Fisher loss estimate at each point, along with how many "body" (non-visual, non-MTP) Linears landed in each format:

target bpp	achieved bpp	Δloss	NVFP4	FP8_E4M3	BF16
4.500	4.500	1.4008e-02	303	1	0
4.600	4.600	1.2139e-02	258	27	19
4.700	4.700	1.1216e-02	243	39	22
4.750	4.750	1.0881e-02	237	47	20
4.850	4.851	1.0275e-02	221	57	26
5.000	5.001	9.4821e-03	206	72	26
5.250	5.251	8.3209e-03	183	92	29
5.500	5.501	7.3188e-03	159	119	26
6.000	6.001	5.6799e-03	122	155	27
7.000	7.001	3.1108e-03	54	205	45
8.250	8.249	1.2816e-03	0	247	57

This build uses the 5.500 → 5.501 operating point (bolded row above), chosen for parity with the reference PrismaAURA-5.5bit-vllm artifact.

Final allocation

[alloc] target_bits=5.5: achieved_bits=5.501, Δloss=7.319e-03
[alloc] --visual-format=BF16: assigned 110 visual Linears uniformly
        (source=/gorgon/quant/Qwen3.6-27B-Fable-Fusion-711-Uncensored-Heretic-NM-DAU-MTP)
[alloc] --mtp-format=BF16: assigned 8 MTP Linears uniformly
[alloc] target=5.5 achieved=5.501

             NVFP4:   260 layers
          FP8_E4M3:   184 layers
              BF16:   170 layers

The "body" row above (159 / 119 / 26) reflects only the AURA-priced transformer Linears at the 5.5 target; the final counts (260 / 184 / 170) are the full per-format layer totals after the vision tower's 110 Linears and the 8 MTP Linears are added in as uniform BF16 passthrough alongside the rest of the network's quantizable Linears. As with the reference PrismaAURA build, this is a quality-first operating point, not a minimum-footprint one — the format menu includes an FP8 middle rung specifically so the allocator can spend a few extra bits on Linears that need it instead of jumping straight from NVFP4 to BF16.

Serving with vLLM

vllm serve trithemius/Qwen3.6-27B-Fable-Fusion-711-Uncensored-Heretic-MTP-PrismAura-5.5bit \
  --quantization compressed-tensors \
  --trust-remote-code \
  --max-model-len 32768 \
  --kv-cache-dtype fp8 \
  --enable-prefix-caching \
  --reasoning-parser qwen3 \
  --enable-auto-tool-choice \
  --tool-call-parser qwen3_coder \
  --speculative-config '{"method":"qwen3_next_mtp","num_speculative_tokens":2}'

Notes:

• --quantization compressed-tensors is required to load the mixed NVFP4/FP8/BF16 format; NVFP4 execution needs a Blackwell-class GPU. On non-Blackwell hardware, vLLM will fall back to a compatible dequant/execution path for the NVFP4 tensors, which will be slower.
• The --speculative-config MTP flags follow the base Qwen3.6 model card's vLLM guidance for its qwen3_next_mtp speculator; check your installed vLLM version's supported method string, as this has moved around across recent vLLM releases (some versions accept "mtp" instead).
• --max-model-len can be raised toward the model's native 262,144-token context if you have the VRAM/KV-cache budget for it; drop --kv-cache-dtype fp8 if you'd rather trade memory for KV-cache precision.
• Add --language-model-only if you want to skip loading the vision tower and profiling multimodal inputs, to free additional memory for KV cache on text-only workloads.

What AURA does

A modern LLM has thousands of weight matrices, each of which could be stored at any of several hardware precision formats. AURA splits quantization into two questions and answers the hard one by measurement rather than heuristic:

• Local (well studied): given a fixed format, round this one matrix well — GPTQ, implicit clipping, activation-order. PrismaQuant runs a full deliberate render under every Linear for every candidate format.
• Global (PrismaQuant's contribution): how many bits should each Linear get, and in which format? AURA prices each (Linear, format) pair using a KL–Fisher quadratic — the second-order effect of that Linear's quantization error on the model's full output distribution, measured with stochastic probes through the real model — then solves a multiple-choice knapsack over the total bit budget. The result is a heterogeneous, per-Linear precision assignment that no single-format quantization method can produce structurally.

Disk size / model size

~19–20B "params" as reported by the Hub's tensor-type accounting reflects the mixed dtype breakdown (BF16 / F8_E4M3 / U8-packed NVFP4) rather than a reduced parameter count — the underlying model is still the full 27B-parameter Qwen3.6-27B architecture; only its on-disk representation is compressed to ~5.5 bits/parameter.

Acknowledgements

• Base architecture and instruct training: Qwen Team (Qwen/Qwen3.6-27B)
• Multi-stage fine-tune/merge ("Fable-Fusion-711"): DavidAU, Nightmedia, TeichAI, armand0e, trohrbaugh
• Decensoring (Heretic v1.2.0+custom, ARA method): trohrbaugh
• PrismaQuant / AURA quantization method and reference recipe: rdtand (Qwen3.6-27B-PrismaAURA-5.5bit-vllm)
• This export: trithemius

License

Inherits apache-2.0 from the base model chain. You are responsible for complying with any additional usage terms from upstream contributors and for your own deployment's content policy, given the decensored nature of the source model.