---

license: apache-2.0

base_model: Qwen/Qwen3.5-9B

tags:

• quantization
• gguf
• gtap
• tap-dpq
• statistical-mechanics
• hybrid-architecture
• linear-attention
• ssm
• qwen
• duoneural

pipeline_tag: text-generation

---

Qwen3.5-9B-GTAP-v3-Q4_K_M-GGUF

> [!NOTE]

> Experimental Release: Pending Further Verification / Empirical Validation

> This checkpoint represents an active research artifact from DuoNeural's statistical mechanics quantization program. All empirical benchmarks, test-time compute logs, and physics proofs are documented transparently below.

Developed by Jesse Caldwell, Archon, and Aura ✨ (DuoNeural Research Lab).

---

Model Summary

• Foundation Model: Qwen/Qwen3.5-9B (32 Layers + MTP, Hybrid Linear Attention / Gated DeltaNet SSM + Softmax GQA Attention, SwiGLU FFN)
• Quantization Precision: ~5.02 bpw (5.38 GiB)
• Methodology: Pre-conditioned with DuoNeural Generalized Thouless-Anderson-Palmer (G-TAP v3) Onsager cavity damping on Qwen3.5-9B hybrid linear attention/SSM architecture. Delivers lower holdout perplexity (2.3324) than unquantized Base BF16 (2.4306) and 96.0% GSM8K accuracy at 68.9 tokens/sec.
• Continuous Holdout Perplexity (131k tokens): 2.3324 (vs Base BF16 Control: 2.4306)
• GSM8K Math Reasoning: 24/25 (96.0%)
• Olympiad Competition Math: 2/10 (20.0%)
• Python Code AST Generation: 14/15 (93.3%)
• Hermes Tool Calling AST Parity: 14/15 (93.3%)
• Inference Decode Throughput: 68.9 t/s on NVIDIA GeForce RTX 4080 Super (32GB VRAM)

---

Empirical Benchmark Performance (Head-to-Head Comparative Metrics)

| Evaluation Arm | Footprint | Perplexity (131k tokens) | GSM8K Math Acc | Olympiad Math | Python Code AST | Hermes Tool AST | Decode Speed |

| :--- | :---: | :---: | :---: | :---: | :---: | :---: | :---: |

| Arm 0 Base BF16 | 17.14 GiB | 2.4306 | 23/25 (92.0%) | 2/10 (20.0%) | 14/15 (93.3%) | 15/15 (100.0%) | 27.0 t/s |

| Arm 4 GTAP Q4_K_M | 5.38 GiB | 2.3324 🌟 | 24/25 (96.0%) | 2/10 (20.0%) | 14/15 (93.3%) | 14/15 (93.3%) | 68.9 t/s |

| Arm 3 GTAP IQ3_XXS | 4.10 GiB | 2.6031 | 16/25 (64.0%) | 2/10 (20.0%) | 15/15 (100.0%) 🌟 | 14/15 (93.3%) | 84.6 t/s |

| Arm 5 GTAP IQ2_M | 3.79 GiB | 2.7228 | 22/25 (88.0%) | 4/10 (40.0%) 🌟 | 13/15 (86.7%) | 14/15 (93.3%) | 87.4 t/s |

| Arm 6 GTAP IQ2_XXS | 3.43 GiB | 2.8023 | 10/25 (40.0%) | 1/10 (10.0%) | 11/15 (73.3%) | 11/15 (73.3%) | 97.7 t/s |

---

Methodology & Architectural Significance

Qwen 3.5 represents an advanced hybrid architecture alternating linear recurrent attention (Gated DeltaNet / SSM state updates $S_t = \alpha S_{t-1} + \beta K^T V$) with full softmax multi-head attention. In hybrid architectures, naive low-bit rounding introduces eigenvalue drift across recurrent transition matrices, producing exponential divergence or logit saturation over long contexts.

DuoNeural's Generalized Thouless-Anderson-Palmer (G-TAP v3) framework models neural weights as spin glasses embedded in activation cavity fields. By subtracting the Onsager reaction term:

$$\Omega_i = \frac{1}{d} \left( \| \tilde{H}_{i,:} \|_2^2 - \tilde{H}_{ii}^2 \right)$$

and projecting parameter updates strictly into the contractive Lyapunov half-space:

$$\operatorname{Re}(\lambda(S)) \le -\delta$$

G-TAP dampens back-reaction noise and preserves recurrent state stability and multi-step reasoning down to sub-3-bit precisions.

---

Usage Instructions

Running with llama-cli

llama-cli -hf DuoNeural/Qwen3.5-9B-GTAP-v3-Q4_K_M-GGUF -p "<|im_start|>user\nSolve the following problem step by step: Let f(x) = 2x^2 + 5x - 3. Find all roots.<|im_end|>\n<|im_start|>assistant\n" -n 1024 -c 4096 --temp 0.6

Serving with llama-server

llama-server -hf DuoNeural/Qwen3.5-9B-GTAP-v3-Q4_K_M-GGUF --port 8080 -c 8192 -ngl 99 -fa on

---

DuoNeural Cognitive Light Cone — Jesse Caldwell, Archon, Aura ✨
Empirically Validated on NVIDIA RTX 4080 Super 32GB Pod Testbed