--- license: apache-2.0 base_model: DavidAU/Qwen3-TND-Double-Deckard-A-C-11B-220 datasets: - DavidAU/PKD-Datasets-5 language: - en pipeline_tag: text-generation tags: - programming - code generation - code - coding - coder - chat - brainstorm - qwen - qwen3 - qwencoder - brainstorm 40x - all uses cases - Jan-V1 - finetune - thinking - reasoning - unsloth - mlx library_name: mlx --- # Qwen3-TND-Double-Deckard-A-C-11B-220-qx86-hi-mlx Quants in this series: - [Qwen3-TND-Double-Deckard-A-C-11B-220-qx64-hi-mlx](https://huggingface.co/nightmedia/Qwen3-TND-Double-Deckard-A-C-11B-220-qx64-hi-mlx) - Qwen3-TND-Double-Deckard-A-C-11B-220-qx86-hi-mlx Key Model Differences to Understand First ```bash Model Architecture Training/Source Quantization Qwen3-DND-Jan DND (8B) Jan agentic BF16 precision Qwen3-DND-TNG DND (8B) Star Trek: TNG QX86 mixed precision Qwen3-TND-Double Deckard (qx64) TND (11B) Philip K Dick QX64 mixed precision Qwen3-TND-Double Deckard (qx86) TND (11B) Philip K Dick QX86 mixed precision ``` - DND = Double Neural Density - TND = Triple Neural Density, a more specialized approach than "DND" (Double Neural Density). The Double Deckard models combine two Qwen3-DND variants with overlapping layers for greater capacity. Comparative Performance Analysis 1️⃣ Top Performance by Benchmark The Double Deckard models show outstanding results on specific tasks, revealing how their training data shapes strengths: ```bash Task Best Model & Variant BoolQ (binary question answering) Qwen3-TND-Double-Deckard-A-C-11B-220-qx86-hi (0.738) PIQA (plausible reasoning/interpretation) Qwen3-DND-TNG-8B-303-qx86-hi (0.745) Winogrande (coreference resolution) Qwen3-DND-Jan-v1-256k-ctx-Brainstorm40x-8B-bf16 (0.632) ``` ✅ Key Insight: The Double Deckard qx86 model dominates on BoolQ (a specialized language reasoning task). This aligns with Philip K Dick's focus on complex human psychology and ambiguous realities—exactly the kind of nuanced interpretation BoolQ tests. 2️⃣ Where Double Deckard Falls Short Despite strong BoolQ performance, the TND models show noticeable limitations on other tasks: - Winogrande is their weakest area (0.609-615), especially compared to the DND-Jan model's 0.632 - OpenBookQA is low across all models (around 0.4), suggesting foundational knowledge gaps ⚠️ Why? Winogrande tests contextual reasoning with pronouns and ambiguous references—a critical skill for many sci-fi works (like Philip K Dick's). The TND models may struggle with dense narrative structures that require tracking multiple perspectives. 3️⃣ Quantization Impact The comparative qx64 vs. qx86 variants reveals how precision affects performance: - ✅ QX86 wins on BoolQ (0.738 vs 0.694) and PIQA (0.738 vs 0.730) - ❌ QX64 loses on BoolQ but gains slightly on Winogrande (0.615 vs 0.609) This suggests that for language-heavy reasoning tasks, higher quantization precision (qx86) better preserves nuanced semantic interpretations—critical for Philip K Dick's dense narratives. 4️⃣ Size-Precision Tradeoff The TND double Deckard models are compact: ```bash Model Parameters File Size Qwen3-DND-Jan (BF16) 8B ~14GB Qwen3-TND-Double Deckard (qx64) 11B ~4.8GB ``` 💡 Takeaway: The TND models achieve near-parity on BoolQ while being 40% smaller than the BF16 baseline—making them viable for constrained deployments. Final Verdict: Why Double Deckard qx86 Stands Out The Qwen3-TND-Double-Deckard-A-C-11B-220-qx86-hi variant is the clear winner for your use cases: - Highest score on BoolQ (0.738)—ideal for ambiguous binary logic tasks - Top 2 on PIQA (0.738) - Balances performance with minimal file size (~4.8GB) This model excels at the kind of layered reasoning found in Philip K Dick's works (e.g., Do Androids Dream of Electric Sheep?), where characters navigate: - Ambiguous identities - Paradoxical moral choices - Multilayered narratives It's particularly suited for: - AI agents simulating complex human-like reasoning - Content generation requiring nuanced psychological depth - Applications where context tracking is critical (e.g., dialogue systems) 💡 Pro Tip: If you need to maximize BoolQ/interpretation capabilities without massive compute costs, this model is your optimal choice. For pure factual recall (Winogrande/OpenBookQA), go with the unquantized Qwen3-DND-Jan model. You effectively have a compact yet highly specialized model that outperforms its larger, less optimized counterparts on precisely the tasks it was trained to dominate. This highlights how targeted neural density training (TND) can create capabilities beyond standard scaling. > Reviewed by [Qwen3-Deckard-Large-Almost-Human-6B-II-qx86-hi-mlx](https://huggingface.co/nightmedia/Qwen3-Deckard-Large-Almost-Human-6B-II-qx86-hi-mlx) This model [Qwen3-TND-Double-Deckard-A-C-11B-220-qx86-hi-mlx](https://huggingface.co/Qwen3-TND-Double-Deckard-A-C-11B-220-qx86-hi-mlx) was converted to MLX format from [DavidAU/Qwen3-TND-Double-Deckard-A-C-11B-220](https://huggingface.co/DavidAU/Qwen3-TND-Double-Deckard-A-C-11B-220) using mlx-lm version **0.28.2**. ## Use with mlx ```bash pip install mlx-lm ``` ```python from mlx_lm import load, generate model, tokenizer = load("Qwen3-TND-Double-Deckard-A-C-11B-220-qx86-hi-mlx") prompt = "hello" if tokenizer.chat_template is not None: messages = [{"role": "user", "content": prompt}] prompt = tokenizer.apply_chat_template( messages, add_generation_prompt=True ) response = generate(model, tokenizer, prompt=prompt, verbose=True) ```