---

license: apache-2.0

base_model: deepseek-ai/DeepSeek-R1-Distill-Qwen-1.5B

tags:

• quantization
• gguf
• gtap
• tap-dpq
• statistical-mechanics
• reasoning
• test-time-compute
• deepseek-r1
• duoneural

pipeline_tag: text-generation

---

DeepSeek-R1-Distill-Qwen-1.5B-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: deepseek-ai/DeepSeek-R1-Distill-Qwen-1.5B (28 Layers, 12:2 GQA, SwiGLU FFN, Test-Time Compute Reasoning)
• Quantization Precision: ~4.5 bpw (1.04 GiB)
• Methodology: Pre-conditioned with Generalized Thouless-Anderson-Palmer (G-TAP v3) Onsager cavity damping. Achieves lower continuous holdout perplexity (4.3641) than the unquantized Base BF16 control model (4.3724) and doubles competition Olympiad math accuracy (40.0% vs 20.0%), confirming that Onsager cavity field subtraction acts as an active thermodynamic noise filter during test-time compute.
• Continuous Holdout Perplexity (131k tokens): 4.3641 (vs Base BF16 Control: 4.3724)
• GSM8K Native CoT Reasoning: 80.0% (20/25)
• Olympiad Competition Math: 40.0% (4/10)
• Inference Decode Throughput: 287.3 t/s on NVIDIA GeForce RTX 4080 Super (32GB VRAM)

---

Empirical Benchmark Performance (Native Test-Time Compute)

| Evaluation Arm | Footprint | Continuous PPL | GSM8K (CoT) | Avg Think (GSM) | Closure (GSM) | Olympiad Math | Decode Speed |

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

| Arm 0 Base BF16 Control | 3.32 GiB | 4.3724 | 20/25 (80.0%) | 409.0 tok | 100.0% | 2/10 (20.0%) | 141.8 t/s |

| Arm 1 Naive IQ3_XXS | 0.72 GiB | 4.7596 | 24/25 (96.0%) | 308.1 tok | 100.0% | 3/10 (30.0%) | 317.9 t/s |

| Arm 3 GTAP v3 IQ3_XXS | 0.72 GiB | 4.7160 | 22/25 (88.0%) | 229.7 tok | 100.0% | 3/10 (30.0%) | 315.0 t/s |

| Arm 4 GTAP v3 Q4_K_M | 1.04 GiB | 4.3641 | 20/25 (80.0%) | 424.1 tok | 100.0% | 4/10 (40.0%) | 287.3 t/s |

| Arm 5 GTAP v3 IQ2_M | 0.65 GiB | 5.1392 | 19/25 (76.0%) | 238.9 tok | 96.0% | 0/10 (0.0%) | 304.9 t/s |

| Arm 6 GTAP v3 IQ2_XXS | 0.55 GiB | 9.4779 | 10/25 (40.0%) | 1250.9 tok | 8.0% | 2/10 (20.0%) | 326.9 t/s |

---

Breakthrough Findings

1. The Super-Perplexity Feat: At Q4_K_M (1.04 GiB), G-TAP v3 achieved a holdout perplexity of 4.3641, strictly lower than the unquantized Base BF16 control model (4.3724). Furthermore, it doubled the competition Olympiad math accuracy (40.0% vs 20.0%) with a 90% clean closure rate.
2. Attractor Regularization in IQ3_XXS: Arm 3 achieved 88.0% GSM8K accuracy with the most concise thinking length (229.7 tokens vs 409.0 on Base BF16), confirming that cavity damping eliminates spurious exploratory branching in reasoning paths.

---

Usage Instructions

Native Prompt Template

DeepSeek-R1 models require the native prompt format to activate the internal chain-of-thought mechanism:

<｜begin of sentence｜><｜User｜>{prompt}<｜Assistant｜><think>

Running with llama-cli

llama-cli -hf DuoNeural/DeepSeek-R1-Distill-Qwen-1.5B-GTAP-v3-Q4_K_M-GGUF -p "<｜begin of sentence｜><｜User｜>Solve: x^2 - 5x + 6 = 0 step by step.<｜Assistant｜><think>\n" -n 1536 -c 4096 --temp 0.6 --top-p 0.95

Serving with llama-server

llama-server -hf DuoNeural/DeepSeek-R1-Distill-Qwen-1.5B-GTAP-v3-Q4_K_M-GGUF --port 8080 -c 4096 -ngl 99 -fa on

---

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