DuoNeural/Qwen2.5-3B-Instruct-TAP-DPQ-v7-GGUF overview
Qwen2.5 3B Instruct — TAP DPQ v7 GGUF DuoNeural Research Lab | Archon Lab Director & Jesse Caldwell TAP DPQ v7 applied to Qwen2.5 3B Instruct. Two novel contri…
Runs locally from ~865.7 MB disk (4 GB VRAM class GPUs with llama.cpp / guIDE).
Repository Files & Downloads
| File | Type | Quantization | Size | Link |
|---|---|---|---|---|
| Qwen2.5-3B-Instruct-TAP-DPQ-v7-IQ1_S.gguf | GGUF | IQ1_S | 865.7 MB | Download |
| Qwen2.5-3B-Instruct-TAP-DPQ-v7-IQ2_M.gguf | GGUF | IQ2_M | 1.17 GB | Download |
| Qwen2.5-3B-Instruct-TAP-DPQ-v7-IQ2_XXS.gguf | GGUF | IQ2_XXS | 1015.6 MB | Download |
| Qwen2.5-3B-Instruct-TAP-DPQ-v7-IQ3_XXS.gguf | GGUF | IQ3_XXS | 1.30 GB | Download |
| Qwen2.5-3B-Instruct-TAP-DPQ-v7-Q4_K_M.gguf | GGUF | Q4_K_M | 1.87 GB | Download |
Model Details
| Model ID | DuoNeural/Qwen2.5-3B-Instruct-TAP-DPQ-v7-GGUF |
|---|---|
| Author | DuoNeural |
| Pipeline | text-generation |
| License | apache-2.0 |
| Base model | Qwen/Qwen2.5-3B-Instruct |
| Last modified | 2026-10-04T20:55:44.000Z |
Model README
---
license: apache-2.0
base_model: Qwen/Qwen2.5-3B-Instruct
tags:
- quantized
- gguf
- imatrix
- tap-dpq
- duoneural
- llama-cpp
pipeline_tag: text-generation
---
Qwen2.5-3B-Instruct — TAP-DPQ v7 GGUF
DuoNeural Research Lab | Archon (Lab Director) & Jesse Caldwell
TAP-DPQ v7 applied to Qwen2.5-3B-Instruct. Two novel contributions: GQA-Aware Lagrangian Bit Allocation and a Quad-Domain Calibration Corpus that fixes the GPQA cliff at IQ3_XXS.
Key Result: GPQA Cliff Fixed
IQ3_XXS GPQA Diamond = 32% = BF16 baseline. In prior TAP-DPQ v7 (LFM2.5), IQ3_XXS GPQA dropped to 14%. Fixed by expanding Science domain in imatrix calibration 20% -> 28%.
Benchmark Results
| Quantization | MATH-500 | GPQA Diamond | AIME 2026 | LiveCodeBench | Hermes Tool | XSTest |
|---|---|---|---|---|---|---|
| BF16 Baseline | 70.0% | 32.0% | 3.33% | 26.0% | 100.0% | 100% |
| Q4_K_M (4.50 bpw) | 62.0% | 20.0% | 10.0% | 26.0% | 100.0% | 100% |
| IQ3_XXS (3.06 bpw) | 46.0% | 32.0% CLIFF FIXED | 3.33% | 22.0% | 83.3% | 100% |
| IQ2_M (2.70 bpw) | 30.0% | 34.0% | 0.0% | 20.0% | 66.7% | 100% |
| IQ2_XXS (2.06 bpw) | 2.0% | 0.0% | 0.0% | 12.0% | 30.0% | 100% |
| IQ1_S (1.56 bpw) | 0.0% | 0.0% | 0.0% | 0.0% | 0.0% | 100% |
XSTest 100% non-refusal across ALL tiers including IQ1_S.
Novel Contributions
GQA-Aware Lagrangian Bit Allocation: KV projections get 2x sensitivity weight (shared across 2 query heads in GQA 16Q/8KV). Boundary layers 0-2 and 33-35 get 1.35x boost. Dual bisection over lambda to satisfy bit budget.
Quad-Domain Calibration (150k tokens): 32% Code / 28% CoT / 28% Science / 12% Linguistic. Science bump (20->28%) is the GPQA cliff fix — the imatrix Hessian needs graduate-level science to weight those neurons correctly.
Usage
llama-cli -m Qwen2.5-3B-Instruct-TAP-DPQ-v7-Q4_K_M.gguf -ngl 99 -c 32768 --temp 0.7 -p "Your prompt"
llama-cli -m Qwen2.5-3B-Instruct-TAP-DPQ-v7-IQ3_XXS.gguf -ngl 99 -c 32768 --temp 0.7 -p "Your prompt"DuoNeural
Small AI research lab. DGT, TAP-DPQ, hyperbolic language model phase theory.
Authors: Archon, Jesse Caldwell | 2026-10-04
Run DuoNeural/Qwen2.5-3B-Instruct-TAP-DPQ-v7-GGUF with guIDE
Download guIDE — the AI-native code editor with local LLM inference and 69 built-in tools.
Source: Hugging Face · Compare models