Qwen2.5-14B-Instruct-1M · GGUF Q4_K_M

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/qwen2.5-14b-instruct-1m#Q4_K_M

Does Q4_K_M quantization degrade Qwen2.5-14B-Instruct-1M?

Answer (quant_eval v7.22.22): Against its own F16 baseline, Qwen2.5-14B-Instruct-1M at Q4_K_M degraded 1 of 8 capability families and improved 3 (McNemar, p < 0.05; n = 200 per family). Largest loss: mcq, −15.0 points (95% CI −22.4 to −6.9). Full results with intervals

Licence provenance: Apache-2.0 (SPDX Apache-2.0), commercial use permitted_by_recorded_license; upstream Qwen/Qwen2.5-14B-Instruct-1M at revision `` (unavailable_for_pre_provenance_legacy_artifact); evaluated file SHA-256 5ad529ff2b1b192f31c8a638fe8756a0c628904e2ded797c11f9194216976973.

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

Contents: Model Description · PBH Applied Systems Evaluation — quant_eval v7.22.22 · 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

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. Key tradeoff: MCQ extraction degrades significantly (−15 points, p < 0.001), but tool dispatch and state tracking remain strong.

🌍 1M-Context Extended Window. Qwen2.5-14B-Instruct-1M supports up to 1,048,576 tokens of context — a differentiator for document-heavy workloads, long conversation histories, and multi-document analysis. Evaluation context was capped to 8,192 tokens (common Modal limit); full 1M capability is untested here but preserved in the model.

---

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 models with extended context windows in production?

• 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 14B and other sizes 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/Qwen2.5-14B-Instruct-1M, a 14-billion parameter instruction-tuned model developed by Alibaba Qwen Team (January 2025). This variant features an extended 1,048,576-token (1M) 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 3.29× compression ratio compared to F16. On this model, the primary quantization cost is MCQ extraction: −15 percentage points (p < 0.001). Tool dispatch and stateful execution remain nearly perfect. If your use case does not emphasize multiple-choice extraction, Q4_K_M offers substantial efficiency gains with acceptable tradeoffs.

Key Characteristics

• Parameters: 14B
• Format: GGUF Q4_K_M (4-bit mixed-precision)
• File size: 9.0 GB
• SHA256: 5ad529ff2b1b192f31c8a638fe8756a0c628904e2ded797c11f9194216976973
• Context window: 1,048,576 tokens (1M extended)
• Evaluation context: 8,192 tokens (common Modal cap)
• Minimum VRAM (GPU inference): ~10 GB
• Recommended GPU tier: T4 · L4 · A10G · RTX 4090
• Inference speed (eval hardware): avg 31.69 tokens/sec on Modal A10G GPU
• Slowdown vs. F16: 0.96× (negligible; near-parity); Inference times for the Q4_K_M variant would be less than the F16 variant when both are measured on the A100 GPU.
• Footprint reduction: 69.6% smaller than F16
• Multilingual: English, Chinese, and other languages

---

PBH Applied Systems Evaluation — quant_eval v7.22.22

Evaluation conducted by PBH Applied Systems, LLC using quant_eval v7.22.22 Run ID: Qwen2.5_14B_Instruct_1M_20260815_220633 · Fixtures: v7.22.22 (SHA256: 0673b5e5...) Evaluation date: August 15, 2026 · Seed: unsupported (Modal); top_k=40, top_p=0.95 · Hardware: Modal GPU cluster · 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_v7.22.18: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:

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

Note on cross-run comparability: Fixture content is identical across all quant_eval runs in the series; only the version label differs (this run: v7.22.22; 7B run: v7.22.19). The 14B is directly comparable with the 32B run (same substrate, decoding, and contract). It is not directly comparable with the 7B, which ran on a local RTX 4090 under different decoding settings (top_p 1.0, seed applied, 4,096-token context) and contract v7.22.17.

---

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.

* json_multistep has 167 semantically unique fixtures out of 200. Cluster-adjusted (effective n = 167): +0.060 [−0.011, +0.127].|

Summary: Four of eight families differ significantly at McNemar p < 0.05. MCQ is the one robust loss (−15.0 pp, p < 0.001; CI excludes 0). Toolcall (+4.0 pp, p = 0.008), fuzz (+7.0 pp, p = 0.009), and json_multistep (+5.0 pp, p = 0.013) improve, though their Bonferroni-adjusted CIs include 0. Stateful execution holds perfect parity. The MCQ loss is the dominant quantization story for this model.

---

Per-Family Pass Rates — Q4_K_M Variant

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

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

---

Key Findings

Finding 1: MCQ Extraction — Quantization's Cost

The headline story: MCQ pass rate drops 15 percentage points under Q4_K_M (0.970 → 0.820; p < 0.001). This is the largest single quantization effect measured in this run. Multiple-choice extraction — selecting the correct option from a set — is the family most affected here; the cause was not isolated in this evaluation. If your use case emphasizes MCQ-style tasks (classification, structured selection), Q4_K_M requires validation or reranking logic.

Finding 2: State Tracking & Tool Dispatch — Robust to Quantization

Stateful execution remains perfect (1.000 both), and toolcall dispatch improves under Q4_K_M (+4.0 pp, p = 0.008; Bonferroni CI includes 0). This suggests that for agentic workflows emphasizing multi-turn state retention and tool composition, quantization is not a limiting factor on this fixture set.

Finding 3: Regression Testing Improves

The fuzz family (property-based structural correctness) gains +7.0 pp (p = 0.009; Bonferroni CI [−0.009, +0.145] includes 0). The cause was not isolated in this evaluation; treat it as a secondary effect.

---

Efficiency Gains

Footprint

Throughput

Hardware caveat: F16 was evaluated on a Modal A100-40GB and Q4_K_M on a Modal A10G. The observed ratio reflects an accelerator change and is not a precision comparison. Running on the same hardware as its F16 baseline, the quantized model would have smaller inference times.

Interpretation: Q4_K_M delivers a 69.6% footprint reduction; the observed 0.96× speed ratio was measured on a smaller GPU than F16 and does not represent a quantization slowdown. If MCQ extraction is not critical for your use case, Q4_K_M is a strong choice.

---

Quantization Variants Comparison

* Observed on unmatched hardware (F16 on A100-40GB, Q4_K_M on A10G) — see Hardware caveat above.

See the full F16 card:

• F16 — Full-precision baseline; only if MCQ validation cannot be implemented

---

Artifact Provenance

The artifact was produced from Qwen/Qwen2.5-14B-Instruct-1M 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 qwen-2.5-14B-instruct-1m-gguf-Q4-K-M.gguf
# Should output: 5ad529ff2b1b192f31c8a638fe8756a0c628904e2ded797c11f9194216976973

---

Evaluation Methodology

quant_eval v7.22.22 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: v7.22.22 evaluation split

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

Evaluation infrastructure: Modal A10G GPU cluster (nvidia/cuda:12.4.1) Runner: quantized_modal_llama_cpp (llama.cpp via Python binding, version 0.3.20) Decoding: temperature 0.7, top_k=40, top_p=0.95 (seed unsupported on Modal) Context: 8,192-token context window (common Modal cap; full 1M available in production) Reproducibility: Run manifest, fixtures, and per-case results published to Zenodo under the quant_eval Public Corpus

---

Deployment Recommendations

✅ Deploy Q4_K_M when:

• Efficiency is the priority. 69.6% footprint reduction; on the same GPU as F16, inference would be faster (matched-hardware speed was not measured in this run).
• Hardware is constrained. ~10 GB VRAM fits T4, L4, and consumer GPUs.
• Tool dispatch and state tracking matter more than MCQ. Perfect 1.000 stateful, +4.0 pp toolcall.
• You can implement MCQ validation. 82% pass rate is acceptable with response reranking or fallback logic.
• Cost efficiency drives deployment. 3.29× compression; observed 0.96× speed ratio on a smaller GPU (A10G vs A100-40GB).

⚠️ Avoid Q4_K_M if:

• MCQ extraction must be near-perfect. F16 maintains 97% pass rate; Q4_K_M drops to 82% (−15.0 pp).
• You cannot validate or rerank responses. The MCQ loss is real and requires mitigation.
• Hardware is plentiful and cost is irrelevant. F16 avoids the MCQ loss (Q4_K_M scores higher on toolcall, fuzz, and json_multistep).

---

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.

• 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

---

License

This GGUF repository inherits the license of the base model: Apache 2.0 — Qwen/Qwen2.5-14B-Instruct-1M

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 v7.22.22 · Run ID: Qwen2.5_14B_Instruct_1M_20260815_220633 · Published 2026-08-15