Qwen3.6 27B Fable Fusion 711 Breaks 0.700 ARC-C on Consumer GPUs

What "711" Actually Means — and Why It Matters

The DavidAU Qwen3.6 27B fine-tune — released in full as Qwen3.6-27B-Fable-Fusion-711-Uncensored-Heretic-NM-DAU-NEO-MAX-MTP-GGUF — has been climbing Hugging Face's trending charts, with the model card and repository listing over 634,000 downloads at the time of writing (a figure that continues to move). The reason for that traction is concrete: according to its authors, it is the first fine-tuned model of its size and type to break an ARC-C score of 0.700 at both 4-bit and 8-bit quantization. Built entirely on consumer hardware (RTX 4090 / RTX 5090 class) using the Unsloth fine-tuning framework, the davidau qwen3.6 27b fable fusion 711 release is described by its authors as a tightly coordinated five-person community effort that pushed a 27-billion-parameter dense transformer well beyond its own base capabilities — and, per the benchmarks in the model card, beyond a larger 35B Mixture-of-Experts sibling in the process.

The number in the model name isn't arbitrary versioning. It points directly to a benchmark milestone reported in the model card: an ARC-C (AI2 Reasoning Challenge) score of 0.711 at mxfp8 precision, and a still-robust 0.701 at mxfp4. The model card is straightforward about it: "The first model of this size/type to breach '700' ARC-C in both 8 bit and 4 bit; hence the '711' in the name."

ARC-C is a widely used proxy for commonsense and scientific reasoning — the kind of multi-step inference that separates a competent assistant from a genuinely useful one. For consumer-grade, locally-runnable open-weights models, clearing 0.700 has historically been a difficult ceiling. That this davidau qwen3.6 fine-tune reportedly achieves it at both quantization tiers suggests the gains are not an artifact of a single precision setting — a distinction that matters for developers choosing quants based on available VRAM. As with any self-reported benchmark, readers should treat these figures as claims from the release team pending independent replication.

Why it matters: When a community fine-tune of a 27B dense model is reported to out-score a larger 35B MoE base model across all seven benchmarks in the model card while remaining fully runnable on consumer GPUs, it challenges the assumption that raw parameter count is the primary driver of local model intelligence.

Architecture at a Glance

Per the model card, the base model is the Qwen3.6-27B-Instruct — described as a 27-billion-parameter dense transformer with a multimodal vision encoder, a 256,000-token context window, and native support for both chain-of-thought (thinking) and direct-answer (non-thinking) inference modes. The fine-tuned variant preserves all of these architectural properties intact; the pipeline adds dataset-driven capability improvements, a decensoring pass, and a specialized quantization stack on top of that foundation. No changes are made to the base model's core weight structure.

The Full Benchmark Picture: Beating the Base Across the Board

According to the model card, all benchmarks were run by collaborator Nightmedia (NM) in instruct mode at both mxfp8 and mxfp4 precision. The results, as reported by the team and compared against the stock base models, are summarized below. These are self-reported figures and have not been independently verified by AIBites.

Model	ARC-C	ARC-E	BoolQ	HellaSwag	OpenbookQA	PIQA	WinoGrande
Fable-Fusion-711 (mxfp8)	0.711	0.879	0.910	0.790	0.514	0.823	0.763
Fable-Fusion-711 (mxfp4)	0.701	0.873	0.909	0.786	0.488	0.813	0.759
Qwen3.6-27B-Instruct (base, mxfp8)	0.647	0.803	0.910	0.773	0.450	0.806	0.742
Qwen3.6-35B-A3B-Instruct (base, mxfp8)	0.581	0.757	0.892	0.751	0.428	0.803	0.688
Qwen3.5-27B-Instruct (base, mxfp8)	0.557	0.711	0.868	0.533	0.452	0.706	0.695

On these reported numbers, the fine-tune beats the base Qwen3.6-27B-Instruct on six out of seven benchmarks and matches it exactly on BoolQ (0.910) — a near-complete sweep. The gains per benchmark are worth spelling out:

• ARC-C: +6.4 percentage points (0.647 → 0.711)
• ARC-E: +7.6 percentage points (0.803 → 0.879)
• HellaSwag: +1.7 percentage points (0.773 → 0.790)
• OpenbookQA: +6.4 percentage points (0.450 → 0.514)
• PIQA: +1.7 percentage points (0.806 → 0.823)
• WinoGrande: +2.1 percentage points (0.742 → 0.763)

These are large margins for post-training fine-tuning, where benchmark movement is normally measured in fractions of a percent — which is itself a reason to treat the numbers cautiously until reproduced. More striking in the model card's comparison is the result against the Qwen3.6-35B-A3B: despite having fewer parameters — 27B dense versus 35B Mixture-of-Experts — Fable-Fusion-711 is reported to beat the larger MoE base on all seven benchmarks. The Qwen3.5-27B comparison in the same table shows a very large reported HellaSwag gap (0.533 → 0.790, over 25 percentage points); a jump that large across a base-model generation is unusual, likely reflects differences in how each base model was evaluated, and should be read as a benchmarker-reported figure rather than a settled result.

Who Built This — and How: A Five-Person Community Pipeline

The long model name functions as a credit roll, and, per the model card, every acronym corresponds to a contributor or methodology. The team describes this as a coordinated, multi-stage pipeline built and validated on consumer hardware rather than a single-person upload:

• DavidAU (DAU): Project lead. Credited with multiple fine-tuning stages using in-house F451-STRICT and Polar-STRICT datasets, and with designing the overall multi-stage merge architecture.
• Nightmedia (NM): Credited with merging logistics and with running the benchmark evaluations at each pipeline stage — the source of the "711" figures cited above.
• TeichAI: Credited with providing the Polaris Dataset — described in the card as GPT-4/GPT-5-derived, non-reasoning synthetic data used to raise general instruction-following quality and response coherence.
• armand0e: Credited with light Fable 5 training traces — creative-writing-oriented data the card associates with the model's prose generation across horror, fiction, roleplay, and mixed-genre storytelling.
• trohrbaugh: Credited with the Heretic decensoring pass using the Arbitrary-Rank Ablation (ARA) method — described below.

The card states the pipeline was built and validated using the Unsloth fine-tuning framework on consumer-grade GPUs — a point the authors emphasize as evidence that this style of model engineering does not require data-center infrastructure. The team also reports first validating the methodology on a smaller Qwen3.5-class 9B model as a test and control group before scaling to the 27B, to reduce the risk of committing expensive compute to an unproven recipe. Per the authors, that smaller validation step is part of why the 27B results are as clean as they are: the dataset mix and merge ordering were refined before scaling.

Decensoring the Refusal Layer: Heretic and ARA Explained

The "Uncensored-Heretic" portion of the name refers, per the model card, to a specific decensoring methodology rather than training on explicit content datasets. The model uses Heretic v1.2.0 + custom, which applies the Arbitrary-Rank Ablation (ARA) technique — a weight-space intervention that aims to remove refusal behaviors by ablating the specific ranks responsible for rejection outputs, without retraining the full model from scratch.

The effect is quantified in the model card as follows:

• KL divergence from the original model: 0.0469 — presented as evidence that the intervention is minimal and the model's core knowledge base is largely undisturbed.
• Refusal rate after ARA: 4 out of 100 test prompts refused — down from 99 out of 100 on the base Qwen3.6-27B-Instruct, per the card's internal test set.

The model card is careful to calibrate expectations: because this is a refusal-removed model rather than one trained on explicit content, default outputs on sensitive topics can trend toward neutral or restrained phrasing. Users after specific tonal registers — slang, explicit language, dark fiction atmosphere — may need to prompt explicitly with descriptors or example phrases to override the model's default neutral register. Per the authors, the base model's instruction-following behavior remains intact; only the refusal gating has been removed. The reported KL divergence of 0.0469 (versus 0.0 for the untouched base) is cited as confirmation that the intervention is precise rather than broad.

The Quantization Stack: NEO, MAX, MTP, and AMD Variants

The model ships with a differentiated set of GGUF quantization options. According to the model card, every quant in the repository — both regular and MTP — is a NEO IMATRIX quant, which the card states improves accuracy by an additional 2–4% over standard GGUFs while also boosting long-context performance. Across all quants, the card notes the output tensor (OT) is stored at full 16-bit precision — a modification the card associates with the MAX designator, intended to improve output fidelity at the top of the logit stack.

Regular vs. MTP Quants

MTP (Multi-Token Prediction) variants use speculative decoding to generate more than one token per forward pass, with MTP tensors stored at Q8_0 precision. Per the card's own testing on an RTX 5090 running LM Studio on Windows 11, Q4_K_S regular GGUFs deliver roughly 75 tokens per second; MTP GGUFs at a 60% token acceptance rate (predicting 2 tokens per pass) can exceed 90 tokens per second — about a 20% speed uplift with no change in output quality. The card notes MTP acceptance rates tend to improve as the context window fills and across multi-turn sessions.

The team specifies a few operational caveats for MTP:

• Keep temperature at or below 1.0 — higher temperatures degrade token acceptance rate and shrink the speed benefit.
• Keep repetition penalty at 1.0 (disabled) — raising it hurts MTP performance.
• If the token acceptance rate drops below 50%, switch to regular (non-MTP) quants for that session.
• For creative writing tasks or temperatures above 1.0, regular GGUFs are the safer choice.

Variant Matrix

Beyond the standard Q4_K_M, Q4_K_S, Q6_K, Q8_0, and IQ4_XS precision tiers, the repository offers two specialized variant families optimized for specific hardware constraints:

Variant Family	Quant Types	Key Difference	Target Use Case
LOW variants	IQ4_XS, Q6_K	Reduced memory footprint	Tight VRAM budgets; lower-end consumer GPUs
AMD/Vulkan variants	IQ4_XS, Q6_K	Output tensor set to f16 instead of bf16	AMD GPUs and Vulkan compute backends; fixes a CPU-offload anomaly on those platforms
MTP variants	All standard tiers	Speculative decoding; MTP tensors at Q8_0	High-throughput inference on RTX 4090 / 5090 class hardware
Standard NEO IMATRIX	Q4_K_M, Q4_K_S, Q6_K, Q8_0, IQ4_XS	OT at 16-bit; NEO IMATRIX calibration	General deployment; maximum compatibility

Capabilities Beyond Benchmarks: Vision, Context, and Creative Writing

Multimodal Vision

Per the model card, the Qwen3.6-27B base is a multimodal architecture, and that vision capability is preserved and activated in this release. Vision processing requires downloading a separate mmproj sidecar file (only one is needed per deployment) and placing it in the same directory as the GGUF. Users running pure text workloads can skip the mmproj file entirely — the text pipeline is unaffected by its absence.

256K Context Window

The model is documented as inheriting the full 256,000-token context window from Qwen3.6-27B. The minimum recommended context allocation at inference time is 8,000–16,000 tokens; NEO IMATRIX quantization is noted in the model card to improve long-context accuracy, which the authors say makes high-context deployments more reliable in quantized form than with standard GGUFs. For developers building document analysis, long-form summarization, or multi-session chat applications, the card's claim is that quantized versions of the davidau qwen3.6 27b model degrade more gracefully at context length than comparable standard GGUF releases.

Thinking and Non-Thinking Inference Modes

Like the base Qwen3.6-27B-Instruct, this model supports both thinking (chain-of-thought reasoning, activated via /think tag or system prompt configuration) and instruct/non-thinking modes (activated via /no_think or by omitting the reasoning prompt). Each mode has its own recommended sampling parameters per the model card:

# Thinking mode — general tasks (reasoning, analysis, Q&A)
temperature=1.0, top_p=0.95, top_k=20,
min_p=0.0, presence_penalty=0.0, repetition_penalty=1.0

# Thinking mode — precise tasks (coding, structured output, WebDev)
temperature=0.6, top_p=0.95, top_k=20,
min_p=0.0, presence_penalty=0.0, repetition_penalty=1.0

# Instruct / non-thinking mode (fast responses, chat, instruction-following)
temperature=0.7, top_p=0.80, top_k=20,
min_p=0.0, presence_penalty=1.5, repetition_penalty=1.0

Per the card, the presence_penalty=1.5 in instruct mode helps prevent repetitive phrasing in direct-answer tasks, while keeping repetition_penalty=1.0 (disabled) stays critical for MTP quant performance regardless of mode.

Creative Writing

Despite its stated primary goal of general intelligence and instruction-following, the Fable 5 training traces from armand0e are credited in the card with giving the model creative range — spanning horror, literary fiction, roleplay, and all-genre storytelling. That combination — multi-step reasoning and expressive, atmospherically controlled fiction — is relatively uncommon in the local open-weights space, where models often excel at one or the other but rarely both. The model card includes sample generations the authors present as tightly paced, immersive prose.

Getting Started: Commands, Hardware Notes, and Deployment Tips

The model is released under an Apache 2.0 license, permitting unrestricted commercial use. It's available directly from the Hugging Face repository under the DavidAU namespace. Pulling and serving a specific quant via llama.cpp's built-in Hugging Face integration is straightforward (binary names may vary slightly by llama.cpp build version):

# Serve via llama-server (compatible with LM Studio, Open WebUI, and similar frontends)
llama-server -hf DavidAU/Qwen3.6-27B-Fable-Fusion-711-Uncensored-Heretic-NM-DAU-NEO-MAX-MTP-GGUF:Q4_K_M

# Or interactive CLI mode
llama-cli -hf DavidAU/Qwen3.6-27B-Fable-Fusion-711-Uncensored-Heretic-NM-DAU-NEO-MAX-MTP-GGUF:Q4_K_M

Key deployment notes:

• Vision workloads: Place the mmproj sidecar file in the same directory as the GGUF and point your inference client to it. Text-only workloads require no mmproj file.
• MTP quants: Identified by an -MTP- suffix in the filename. Drop-in compatible with standard GGUFs; the only runtime difference is a potential inference speed gain. Requires temperature ≤ 1.0 for reliable acceptance rates.
• AMD / Vulkan users: Use the dedicated AMD/Vulkan variant files (output tensor set to f16) to avoid CPU-offload anomalies.
• Context allocation: Set a minimum context of 8,000–16,000 tokens at inference time; under-allocating context is, per the card, a common cause of degraded output quality at scale.
• Hardware baseline: Q4_K_S fits comfortably on a 24 GB VRAM GPU (RTX 3090/4090/5090); Q8_0 typically requires 32 GB+ VRAM or CPU offload. The quoted speed figures (75–90+ tokens/sec) are from the card's own RTX 5090 testing in LM Studio on Windows 11.

Key Takeaways

• Reported first 27B fine-tune to clear 0.700 ARC-C at both 4-bit and 8-bit: The "711" name refers to model-card benchmarks by Nightmedia of 0.711 (mxfp8) and 0.701 (mxfp4) — self-reported and awaiting independent replication.
• Reported to beat a larger base model across all seven benchmarks: Per the card, the davidau qwen3.6 27b fable fusion 711 outperforms the Qwen3.6-35B-A3B-Instruct MoE base on every evaluated benchmark despite having fewer parameters.
• Community-built on consumer hardware: Five contributors, the Unsloth framework, and RTX 4090/5090-class GPUs — with the authors reporting a smaller 9B validation run before scaling to 27B.
• ARA-based decensoring is presented as surgical, not destructive: A reported KL divergence of 0.0469 from the base; refusal rate reportedly drops from 99% to 4% on the card's internal test set.
• MTP quants reportedly offer ~20% inference speed uplift (card-tested at 90+ tokens/sec on RTX 5090 vs. ~75 for regular quants), with no functional difference in output — but require temperature ≤ 1.0 and repetition penalty disabled.
• Full 256K context window preserved and, per the card, improved for long-context accuracy via NEO IMATRIX quantization across all released quant tiers.
• Multimodal vision is active and requires only an additional mmproj sidecar file — no architectural changes needed.
• Apache 2.0 licensed: Commercial use is permitted without restriction.

---

If the model card's figures hold up under independent testing, this release points toward a maturing community fine-tuning ecosystem — one where small, coordinated teams with targeted datasets and well-ordered merge strategies can extract performance that rivals larger models. The davidau qwen3.6 methodology the team describes — validate on a smaller model, then scale to 27B — reads as a reusable, low-waste playbook for the open-source community. It is reasonable to expect similar multi-stage pipelines applied to other base models, and for the broader community to build on both the Fable creative-writing trace approach and the Heretic ARA decensoring technique as open, composable primitives. With the repository actively trending on Hugging Face and download counts in the hundreds of thousands at the time of writing, the appetite for principled, benchmark-driven fine-tuning work is clearly there — pending the independent replication that would turn these self-reported results into an established benchmark.