---

language:

• en
• zh

license: apache-2.0

base_model: Qwen/Qwen2.5-7B-Instruct

tags:

• quant_eval
• pbh-applied-systems
• gguf
• q4_k_m
• quantization
• qwen
• instruct
• llama-cpp
• agentic
• tool-calling
• structured-output
• multilingual
• quantization-degradation
• behavioral-evaluation
• llm-evaluation

model-index:

• name: qwen-2.5-7B-instruct-gguf-Q4-K-M

results:

- task:

type: text-generation

name: Agentic behavioral evaluation (quant_eval)

dataset:

name: quant_eval golden oracle fixtures (v7.22.21)

type: pbhappliedsystems/quant_eval_golden_oracle_fixtures

metrics:

- type: pass_rate

name: json pass rate

value: 38.0

- type: pass_rate

name: json_multistep pass rate

value: 54.0

- type: pass_rate

name: stateful_followup pass rate

value: 87.0

- type: pass_rate

name: mixed_brief_json pass rate

value: 92.5

- type: pass_rate

name: mcq pass rate

value: 84.5

- type: pass_rate

name: toolcall pass rate

value: 72.5

- type: pass_rate

name: toolcall_only pass rate

value: 82.0

- type: pass_rate

name: fuzz pass rate

value: 50.5

source:

name: quant_eval Public Corpus

url: https://doi.org/10.5281/zenodo.22009419

---

Qwen2.5-7B-Instruct · GGUF Q4_K_M

<!-- pbh-site-link -->

> Evaluation results: this Q4_K_M build compared with F16, capability family by family, with confidence intervals and file size, on the results page: pbhappliedsystems.com/models/qwen2.5-7b-instruct#Q4_K_M

<!-- pbh-card4 -->

Does Q4_K_M quantization degrade Qwen2.5-7B-Instruct?

Answer (quant_eval v7.22.21): Against its own F16 baseline, Qwen2.5-7B-Instruct at Q4_K_M degraded 2 of 8 capability families and improved 2 (McNemar, p < 0.05; n = 200 per family). Largest loss: mixed_brief_json, −7.5 points (95% CI −12.8 to −1.8). Full results with intervals

!Qwen2.5-7B-Instruct: Q4_K_M, 2 of 8 families degraded, 2 improved

Licence provenance: Apache-2.0 (SPDX Apache-2.0), commercial use permitted_by_recorded_license; upstream Qwen/Qwen2.5-7B-Instruct at revision ` (unavailable_for_pre_provenance_legacy_artifact); evaluated file SHA-256 863656d217841f5d3fb180d9dca4e4bbdaa071bde25885fa0d27fe7188a2cc85`.

Other builds of this model: F16 · All evaluated builds (Collection) · Datasets

Contents: Model Description · PBH Applied Systems Evaluation — quant_eval v7.22.21 · Comparability & Fixture Methodology · Paired Comparison: Q4_K_M vs. F16 Baseline · Per-Family Pass Rates — Q4_K_M Variant · Key Findings · Efficiency Gains · Quantization Variants Comparison

<!-- /pbh-card4 -->

Converted and evaluated by PBH Applied Systems, LLC

— Applied AI/ML Consulting · LLM Optimization & Deployment · Quantized AI Infrastructure

> 🔬 This repository is part of a production-oriented evaluation series. Every model published under pbhappliedsystems has been independently evaluated using quant_eval v7.22.21/v7.22.22 — a behavioral evaluation harness developed by PBH Applied Systems. Scores measure real agent-adjacent task performance across structured output, tool dispatch, multi-turn state retention, and multi-step planning families — not perplexity or benchmark leaderboard proxies.

> 📌 This is the Q4_K_M quantized variant. The full-precision F16 baseline is published separately: F16. This card shows how Q4_K_M compression compares to F16 across all behavioral dimensions. For many production use cases, Q4_K_M offers the better price–performance tradeoff — with surprising behavioral gains on planning tasks.

---

Try the Live AI Agent Demo & Compare Models

Two complementary interfaces, both free:

1. Quick Demo — Single Model Testing

Launch the PBH Applied Systems AI Agent Demo →

Free · 3 example queries per agent type · No account required · No document upload

Test this model (or others) against three agent workflow templates:

• Reasoning & Analysis — Chain-of-thought decomposition and structured problem-solving

Example queries: Should a startup build on cloud LLMs or self-host quantized models? · Analyze the trade-offs between model quantization and inference latency. · What are the cost implications of running 7B parameter models on edge devices?

• Document Intelligence — Long-context information extraction and Q&A

Example queries: Extract key clauses from a contract and summarize legal risks. · Analyze a research paper and identify the novel contributions. · Compare market analysis reports and highlight strategic differences.

• Code & Automation — API scaffolding, data transformation, and task completion

Example queries: Build a REST API endpoint for user authentication and include rate limiting. · Transform CSV data into a production-ready database schema. · Write a Python script to validate and clean messy customer records.

Every query runs on private GPU infrastructure — your prompts, reasoning traces, and outputs never leave the system and are never sent to Frontier model providers or cloud vendors. This matters for organizations with sensitive data, compliance requirements, or internal knowledge that cannot leave the building.

2. Model Comparison Arena — Side-by-Side Evaluation

Launch the quant_eval Agent Arena Space →

Compare two models in real-time across all three agent types.

The Agent Arena lets you:

• Select any two evaluated models (F16 or quantized variants) and test them side-by-side on the same queries
• View execution traces for both agents, showing chain-of-thought reasoning and tool dispatch
• Check the Model Leaderboard — the Q4_K_M pass rate in each of the eight quant_eval capability families for every model with a published run, marked ▼ degraded or ▲ improved where McNemar's paired test against the model's own full-weight baseline is significant, ordered by how many families degraded
• Read the Methodology tab — explanation of quant_eval, the 8 test families, and how to request a full evaluation report for models not yet in the series

Why compare in the Arena:

The demo shows what this model can do. The Arena shows how this model compares to others. If you're deciding between F16 and Q4_K_M, or between this model and another in the series, run the same query in the Arena with both selected. You'll see the exact differences in reasoning quality, tool dispatch, and output coherence — not benchmark scores, but real agent behavior.

Evaluation-backed leaderboard: Every score in the Arena comes from published quant_eval results: the quant_eval Public Corpus (DOI 10.5281/zenodo.22009419) and, for showcase models such as Ministral-3-14B-Instruct-2512, sealed run bundles published as summaries. The numbers are not opinionated; they're measured.

---

Model Description

This repository contains the Q4_K_M quantized GGUF of Qwen/Qwen2.5-7B-Instruct, a 7-billion parameter instruction-tuned model developed by Alibaba Qwen Team (2024). Qwen2.5-7B-Instruct features a 32,768-token context window and multilingual instruction-following capability.

Q4_K_M is a mixed-precision K-quant: mostly 4-bit (Q4_K) weights, with selected tensors kept at 6-bit (Q6_K), delivering a 3.25× compression ratio compared to F16. On this model, Q4_K_M shows counterintuitive behavioral gains on multi-step planning (json_multistep: +18.5 percentage points) and stateful conversation (+12.0 pp), no significant change on tool dispatch, and significant losses on hybrid responses (−7.5 pp) and single-step JSON (−4.5 pp, p = 0.049).

For most production deployments requiring speed and efficiency, Q4_K_M is the recommended choice.

Key Characteristics

• Parameters: 7B
• Format: GGUF Q4_K_M (4-bit mixed-precision)
• File size: 4.68 GB
• SHA256: 863656d217841f5d3fb180d9dca4e4bbdaa071bde25885fa0d27fe7188a2cc85
• Context window: 32,768 tokens
• Minimum VRAM (GPU inference): ~7 GB
• Recommended GPU tier: RTX 4090 · T4 · L4 · Mobile GPU (16 GB / 8 GB)
• Inference speed (eval hardware): avg 119.12 tokens/sec on NVIDIA RTX 4090
• Speedup vs. F16: 2.15× faster (wall-time: 1.98×)
• Footprint reduction: 69.3% smaller than F16
• Multilingual: English, Chinese, and other languages

---

PBH Applied Systems Evaluation — quant_eval v7.22.21

> Evaluation conducted by PBH Applied Systems, LLC using quant_eval v7.22.21

> Run ID: Qwen2.5_7B_Instruct_20260814_234822 · Fixtures: v7.22.19 (SHA256: 2a0f7cdd...)

> Evaluation date: August 14, 2026 · Seed: 42 · Hardware: NVIDIA RTX 4090 · Published: August 15, 2026

Evaluation Data Published to Zenodo

All evaluation artifacts are published under the quant_eval Public Corpus

(DOI: 10.5281/zenodo.22009419):

• Run Provenance (D4): 10.5281/zenodo.22010462 — Selected run manifest and rollup fields, artifact SHA-256 hashes, runner and decoding configurations
• Golden Oracle Fixtures (D3): 10.5281/zenodo.22010278 — Test case specifications and fixture version crosswalk
• Family Pass Rates (D6): 10.5281/zenodo.22010623 — Aggregate pass rates per family, both runners, 95% Wilson confidence intervals
• Per-Case Behavioral Results (D1): 10.5281/zenodo.22009799 — 19,200 per-case rows, both precisions, all signals, case-level pass/fail
• Throughput Telemetry (D2): 10.5281/zenodo.22009987 — Per-call generation metrics, wall-time, token counts
• Efficiency & Footprint (D7): 10.5281/zenodo.22010723 — Stored artifact size, compression ratio, tokens/sec, observed wall-time ratios with hardware scope

Evaluation Contract: canonical_agentic_contract_v7.22.17:evaluation

---

Comparability & Fixture Methodology

Fixtures Enable Valid Variant Comparison

The quant_eval evaluation harness uses a consistent fixture set to measure both full-precision and quantized variants. The F16 and Q4_K_M variants were evaluated under the same fixture set, enabling direct comparison.

Fixture Generation:

| Variant | Run ID | Fixture SHA256 | Evaluation Contract | Rationale |

|---|---|---|---|---|

| F16 & Q4_K_M | 20260814_234822 | 2a0f7cdd... | v7.22.17:evaluation | Single run evaluating both runners sequentially |

What this means: Results are directly comparable. Both runners used identical fixtures, enabling precise measurement of Q4_K_M behavioral change relative to F16.

---

Paired Comparison: Q4_K_M vs. F16 Baseline

Quantization effect on all families — directly measured.

Deltas represent pass-rate change (Q4_K_M − F16), with 95% confidence intervals (paired multinomial, Bonferroni-adjusted Wilson) and paired McNemar p-values for statistical significance. N = 200 cases per family.

| Family | F16 Rate | Q4_K_M Rate | Δ | 95% CI | McNemar p | Pattern |

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

| json_multistep | 0.355 | 0.540 | +0.185 | [+0.098, +0.263] | 7.84e-10 | Significant gain |

| stateful_followup | 0.750 | 0.870 | +0.120 | [+0.053, +0.181] | 1.19e-07 | Significant gain |

| toolcall_only | 0.805 | 0.820 | +0.015 | [−0.027, +0.056] | 0.375 | Nominal |

| mixed_brief_json | 1.000 | 0.925 | −0.075 | [−0.128, −0.018] | 6.10e-05 | Significant loss |

| json | 0.425 | 0.380 | −0.045 | [−0.109, +0.021] | 0.049 | Borderline loss (p < 0.05; CI includes 0) |

| fuzz | 0.520 | 0.505 | −0.015 | [−0.085, +0.055] | 0.648 | Nominal |

| toolcall | 0.710 | 0.725 | +0.015 | [−0.071, +0.101] | 0.720 | Nominal |

| mcq | 0.835 | 0.845 | +0.010 | [−0.029, +0.049] | 0.625 | Nominal |

Adjusted for semantic clustering (json_multistep): 167/200 semantically unique cases; cluster-adjusted delta = +0.186 [+0.088, +0.272], effective n=167. The reported gain remains statistically significant.

Summary: Q4_K_M delivers +0.185 improvement on multi-step planning and +0.120 on stateful conversation — counterintuitive gains that merit investigation. The hybrid family (mixed_brief_json) is the largest loss (−0.075, p < 0.001), dropping from perfect to 92.5%; single-step json shows a borderline loss (−0.045, p = 0.049).

---

Per-Family Pass Rates — Q4_K_M Variant

All rates include 95% Wilson score confidence intervals. N = 200 cases per family.

| Family | Pass Rate | 95% CI | Notes |

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

| mixed_brief_json | 0.925 | [0.880, 0.954] | Hybrid answer+JSON; one in thirteen fails under quantization |

| stateful_followup | 0.870 | [0.816, 0.910] | +12.0 pp over F16; strong state tracking |

| toolcall_only | 0.820 | [0.761, 0.867] | +1.5 pp over F16 (not significant) |

| toolcall | 0.725 | [0.659, 0.782] | End-to-end tool dispatch |

| mcq | 0.845 | [0.788, 0.889] | +1.0 pp over F16 (not significant) |

| json | 0.380 | [0.316, 0.449] | −4.5 pp compared to F16 |

| fuzz | 0.505 | [0.436, 0.574] | Regression: property-based checks |

| json_multistep | 0.540 | [0.471, 0.608] | +18.5 pp over F16; planning pass rate improves under quantization |

Source: quant_eval Family Pass Rates, DOI 10.5281/zenodo.22010623

---

Key Findings

Finding 1: Q4_K_M Outperforms F16 on Multi-Step Planning

The most striking result: json_multistep passes improve 18.5 percentage points when using Q4_K_M (0.355 → 0.540; p < 0.001), and the gain holds under cluster adjustment. The cause of this reversal was not isolated in this evaluation and warrants further investigation. On this fixture set, Q4_K_M scored higher than F16 on multi-step planning.

Finding 2: State Tracking — 87% Success with Q4_K_M

Stateful followup improves significantly (+12.0 pp, p < 0.001), reaching 87% pass rate. This is enterprise-usable: 13 failures per 100 two-turn conversations is acceptable in contexts where retry logic or validation is available. F16 at 75% leaves more room for failure.

Finding 3: Hybrid Responses — The Largest Measured Tradeoff

The largest loss is in hybrid answers (−7.5 pp, p < 0.001): mixed_brief_json drops from perfect (1.000) to 0.925. This family requires both natural language and JSON to be correct simultaneously. The loss is real but bounded: 185 of 200 cases still pass, meaning failures are sparse and potentially correctable via response reranking.

---

Efficiency Gains

Footprint

| Metric | F16 | Q4_K_M | Reduction |

|---|---|---|---|

| File size | 15.2 GB | 4.68 GB | −69.3% |

| GPU VRAM needed | ~17 GB | ~7 GB | −59% |

| Disk I/O (during loading) | 15.2 GB | 4.68 GB | −69.3% |

Throughput

| Metric | F16 | Q4_K_M | Speedup |

|---|---|---|---|

| Tokens/sec | 55.44 | 119.12 | 2.15× |

| Generation time per 1K tokens | 18.0 sec | 8.4 sec | 2.15× faster |

| Evaluation wall time (1,600 cases) | 1,479 sec | 745 sec | 1.98× faster |

Hardware: NVIDIA RTX 4090 (same GPU for both precisions), llama.cpp backend, 4K context window

Interpretation: Q4_K_M is substantially faster — a 2.15× generation throughput increase and a 1.98× observed evaluation wall-time ratio on the same hardware. Combined with its planning gains, this makes Q4_K_M a strong fit for latency-sensitive deployments that can tolerate the hybrid-response loss.

---

Quantization Variants Comparison

| Variant | File Size | VRAM | Speed | Stateful | Toolcall (end-to-end) | Hybrid | Planning | Best For |

|---|---|---|---|---|---|---|---|---|

| F16 | 15.2 GB | ~17 GB | 1.0× | 0.750 | 0.710 | 1.000 | 0.355 | Maximum fidelity; hybrid-response workloads |

| Q4_K_M | 4.68 GB | ~7 GB | 1.98× | 0.870 | 0.725 | 0.925 | 0.540 | Recommended for most production |

How to read this table: Speed is the observed evaluation wall-time ratio (F16 ÷ quantized) on the same RTX 4090 — not a controlled throughput benchmark. Generation throughput ratio for the same pair is 2.15×.

See the full F16 card:

• F16 — Full-precision baseline; when maximum fidelity is non-negotiable

---

Artifact Provenance

| Artifact | Format | Size | SHA256 |

|---|---|---|---|

| Qwen2.5_7B_Instruct_Q4_K_M.gguf | GGUF Q4_K_M | 4.68 GB | 863656d217841f5d3fb180d9dca4e4bbdaa071bde25885fa0d27fe7188a2cc85 |

The artifact was produced from Qwen/Qwen2.5-7B-Instruct using a custom-built llama.cpp conversion and quantization pipeline developed by PBH Applied Systems.

Artifact verification: SHA256 hashes are recorded in the run manifest (DOI 10.5281/zenodo.22010462). Download the Q4_K_M GGUF and verify:

sha256sum Qwen2.5_7B_Instruct_Q4_K_M.gguf
# Should output: 863656d217841f5d3fb180d9dca4e4bbdaa071bde25885fa0d27fe7188a2cc85

---

Evaluation Methodology

quant_eval v7.22.21 is a behavioral evaluation harness developed by PBH Applied Systems. The two-run architecture evaluates the full-precision (F16) model first, then evaluates the quantized variant against the same fixture set, enabling direct measurement of behavioral degradation or improvement.

Fixture set: v7.22.19 evaluation split

• Total unique cases: 1,600 (200 per family × 8 families)
• Per-family test families and pass signals:

| Family | Description | Pass Signals |

|---|---|---|

| json_multistep | Multi-step planning with self-check and oracle verification | parse_ok, oracle_equiv_ok |

| stateful_followup | Two-turn state tracking; turn-2 correct given turn-1 | turn1/2_parse_ok, turn1/2_exact_match |

| toolcall_only | Bare schema-only tool call; strict tool name + args check | tool_name_ok, args_ok |

| mixed_brief_json | Hybrid: natural language answer + valid JSON block | answer_line_ok, json_parse_ok, schema_ok, semantic_payload_ok |

| json | Single-step structured JSON with constraint rules | schema_ok, final_equiv_ok, constraints_ok |

| fuzz | Property-based regression; structured placement correctness | plan_exact_match, final_equiv_ok, constraints_ok |

| toolcall | Tool call embedded in response; parse + schema validation | stage1_tool_parse_ok, selection_ok, stage1_args_ok, tool_exec_ok, stage2_evaluable, final_equiv_ok |

| mcq | Multiple-choice extraction | semantic_choice_ok |

Evaluation hardware: NVIDIA RTX 4090 (24 GB VRAM)

Runner: quantized_llama_cpp (llama.cpp via Python binding, version 0.3.16)

Decoding: temperature 0.7, seed 42, top_p 1.0

Context: 4,096-token context window (llama.cpp configuration)

Reproducibility: Run manifest, fixtures, and per-case results published to Zenodo under the quant_eval Public Corpus

---

Deployment Recommendations

✅ Deploy Q4_K_M when:

• Speed and efficiency are critical. 2.15× generation throughput (1.98× observed wall-time) on the same hardware, with controlled tradeoffs.
• Hardware is constrained. ~7 GB VRAM fits T4, L4, mobile GPU, and single consumer cards.
• Multi-step planning is important. Q4_K_M outperforms F16 on json_multistep (+18.5 pp, p < 0.001).
• You have retry/validation logic. The hybrid loss (−7.5 pp) is acceptable with response reranking or validation.
• Cost efficiency matters. 69% footprint reduction, dramatically cheaper inference.

⚠️ Avoid Q4_K_M if:

• Hybrid responses must be perfect. F16 maintains 100% pass rate; Q4_K_M is 92.5%.
• You're latency-indifferent and VRAM-unlimited. F16 preserves maximum fidelity at the cost of speed.
• Quantization noise is unacceptable for your domain. Some edge cases may introduce variability.

---

About PBH Applied Systems

PBH Applied Systems, LLC is an Oklahoma City–based applied machine learning and AI systems company specializing in production-grade model evaluation, quantization pipelines, agentic AI infrastructure, and scalable AI-driven application development. The organization emphasizes engineering rigor, reproducibility, and real-world deployment constraints — particularly in environments where performance, cost efficiency, and reliability must be balanced against available hardware and budget.

Founder & Principal AI/ML Systems Architect — Patrick Hill

Patrick Hill is the Founder & Principal AI/ML Systems Architect of PBH Applied Systems with 10+ years of experience delivering advanced analytics, predictive modeling, and decision-support solutions across high-stakes operational environments. Patrick holds a Master of Science in Software Engineering with concentrations in Artificial Intelligence and Machine Learning and a B.S. in Business Finance.

Technical expertise spans: Python, SQL, Linux, Pandas, NumPy, scikit-learn, PyTorch, TensorFlow/Keras, HuggingFace Transformers, GGUF, llama.cpp, BitsAndBytes, PEFT, QLoRA, Flask APIs, Docker, CI/CD, Jupyter, Databricks, and quantization strategies.

Published Author: Patrick is the author of Applied Machine Learning: Concepts, Tools, and Case Studies — a 1,200+ page practitioner-oriented textbook adopted as required reading for CSC 373 – Machine Learning at the University of Advancing Technology.

Core Service Areas: LLM optimization & deployment · AI evaluation frameworks · Agentic AI infrastructure · Scalable AI application development · ML pipeline design & analytics · Model & agent cataloging.

---

🔬 About quant_eval & This Evaluation Series

quant_eval is a behavioral evaluation harness developed by PBH Applied Systems, LLC. It measures real agent-adjacent task performance across structured output, tool dispatch, multi-turn state retention, and multi-step planning — not perplexity or leaderboard proxies. Every model published under pbhappliedsystems has been independently evaluated using quant_eval before being recommended for any production role.

See it in action: Live AI Agent Demo →

---

📋 Interested in Deeper Engagement?

This model card documents behavioral evidence under quant_eval. If you're evaluating this model for production deployment, need guidance on quantization impact for your specific workload, or want to understand deployment tradeoffs under your infrastructure constraints, PBH Applied Systems offers several paths:

Async Intake Form — https://pbhappliedsystems.com/contact

Describe your use case, infrastructure, and evaluation needs. Responses are reviewed asynchronously.

Available Services:

• Evaluation Report — A written behavioral audit: per-family pass rates, F16 vs. quantized delta analysis, failure cluster diagnostics, deployment recommendation, and a quantization-level suggestion for your hardware and latency constraints. Evidence-standard numbers only. Typical engagements: $2,500–$5,000.

• Sample report — Read a complete sample report: a real evaluation report for Ministral-3-14B-Instruct-2512, its BF16 baseline against Q4_K_M, Q5_K_M and Q8_0, written exactly as a client receives it. <!-- pbh-sample-report -->

• Starter-Kit Recommender — Configured agent templates (reasoning, document intelligence, code automation) with proof metrics from your specific workload. Includes model selection, quantization strategy, and deployment architecture guidance tied to your infrastructure.

Submit your details via the form above, and we'll route you to the appropriate engagement.

---

💬 Community

Questions about this model, a model or quant you want measured next, or your own quantization results to compare:

• Discord — PBH Applied Systems: #model-requests, #your-evals, #deployment-help
• GitHub Discussions — quant_eval Public Corpus tools: model requests and polls on what to measure next
• This repo — the Community tab for anything specific to these files

<!-- pbh-community -->

---

License

This GGUF repository inherits the license of the base model:

Apache 2.0 — Qwen/Qwen2.5-7B-Instruct

The quant_eval evaluation harness, fuzz prompt builder, and scoring implementation are proprietary to PBH Applied Systems, LLC and are not included in this repository. The golden oracle fixture set used in this evaluation is published under CC BY 4.0 as part of the quant_eval Public Corpus (D3, DOI 10.5281/zenodo.22010278).

---

GGUF quantization and behavioral evaluation performed by PBH Applied Systems, LLC · quant_eval v7.22.21 · Run ID: Qwen2.5_7B_Instruct_20260814_234822 · Published 2026-08-15