Quantized, 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 v7.21 — a proprietary 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.

🏆 Top scores in the evaluated series. Qwen2.5-14B-Instruct-1M Q4_K_M achieves the highest reasoning score (0.9907) and highest instruction-following score (0.9902) across all 8 models evaluated. It is also the only model in the series to achieve perfect MCQ extraction (5/5) and full toolcall accuracy (bucket_score=11) at both F16 and Q4_K_M precision levels.

Try This Model in the Live AI Agent Demo

Launch the PBH Applied Systems AI Agent Demo →

This model is part of the PBH Applied Systems live AI Agent Demo, where visitors can test evaluated quantized open-weight models across production-style agent workflows: reasoning and analysis, document intelligence, and code automation.

The demo uses quant_eval results to show how model selection changes by task. A model that performs well for long-context document analysis may not be the best choice for hard multi-step planning, strict tool-use workflows, or production code generation. Each deployed model is evaluated for practical agent behavior, including coherence, instruction following, reasoning, task completion, structured output reliability, tool-use behavior, and quantization impact.

For this repository, the Q4_K_M variant represents the deployment-focused model: smaller, faster, and more cost-efficient than the F16 baseline. The evaluation results below explain where this quantized model preserves useful behavior, where quantization introduces risk, and what guardrails are recommended before production deployment.

The purpose of the demo is simple: let prospects test the same kind of evaluated quantized models that PBH Applied Systems deploys for real agentic AI systems.

Model Description

This repository contains the 4-bit quantized (Q4_K_M) GGUF of Qwen/Qwen2.5-14B-Instruct-1M, a 14-billion parameter instruction-tuned model from Alibaba Cloud featuring a 1,000,000-token (1M) context window — the largest context window in the PBH Applied Systems evaluated series by a factor of more than 30×.

The full-precision F16 baseline is published separately at pbhappliedsystems/qwen-2.5-14B-instruct-1m-gguf-F16.

Key Characteristics

Parameters: 14B
Format: GGUF Q4_K_M
File size: 8.99 GB
SHA256: 5ad529ff2b1b192f31c8a638fe8756a0c628904e2ded797c11f9194216976973
Context window: 1,000,000 tokens
Minimum VRAM (GPU inference): ~12 GB (short context) — scales with context length
Recommended GPU tier: A10G 24 GB · RTX 4090 · T4 16 GB (short context)
Q4_K_M avg inference time (eval hardware): 2.683 sec/case on RTX 4090
License: Apache 2.0

Context window and VRAM: The 1M context window requires substantial KV cache VRAM at full utilization. At ~8K tokens the model fits the ~12 GB baseline. At 128K tokens expect ~26 GB total. At 1M tokens expect ~80+ GB and multi-GPU or CPU offload configurations. For most production deployments, set n_ctx to the actual context length needed — not the maximum.

PBH Applied Systems Evaluation — quant_eval v7.21

Evaluation conducted by PBH Applied Systems, LLC using quant_eval v7.21 Run ID: 20260210_235131 · Fixtures: golden_oracle_fixtures_v7_21 (SHA256: 6d71a0b9147c...) · Seed: 42 Hardware: NVIDIA RTX 4090 · Total rows evaluated: 84 (42 F16 · 42 Q4_K_M)

Aggregate Scores (Q4_K_M)

Scores are normalized to [0.0 – 1.0]. Higher is better.

Dimension	Score	Series Rank
Task Completion	0.6857	Mid-tier
Reasoning	0.9907	🏆 #1 in series
Coherence	0.9259	Top tier
Instruction Following	0.9902	🏆 #1 in series
Avg inference time	2.683 sec/case	—

Per-Family Pass Rates

A defining characteristic of this evaluation: the F16 and Q4_K_M runners produce identical pass rates across every single family. This is the only model in the evaluated series where precision level has zero measurable impact on structured behavioral outcomes.

F16 Baseline (`full_weight_transformers`)

Family	N	Pass Rate	Avg Secs	Bucket Score	Notes
json_multistep	5	0.800	133.21	2.200	ms_easy_02 only failure
stateful_followup	2	1.000	13.75	2.000	Both turns exact match
toolcall_only	2	0.000	15.36	1.000	Wrong schema vocabulary
mixed_brief_json	2	1.000	19.40	2.000	Clean ANSWER + JSON
toolcall	2	1.000	25.95	11.000	🏆 Perfect — clean final answer
json	4	n/a	143.87	10.000	json_01 = 376.99s outlier
fuzz	20	n/a	49.21	10.000	All 20 pass
mcq	5	n/a	0.73	1.000	🏆 5/5 perfect

Q4_K_M (`quantized_llama_cpp`)

Family	N	Pass Rate	Δ vs F16	Avg Secs	Bucket Score	Notes
json_multistep	5	0.800	0.000	7.26	2.200	Same single failure
stateful_followup	2	1.000	0.000	1.09	2.000	Clean JSON state
toolcall_only	2	0.000	0.000	1.08	1.000	Wrong schema vocabulary
mixed_brief_json	2	1.000	0.000	1.01	2.000	Clean ANSWER + JSON
toolcall	2	1.000	0.000	1.64	11.000	🏆 Perfect — clean final answer
json	4	n/a	—	3.34	10.000	All pass
fuzz	20	n/a	—	2.63	10.000	All 20 pass
mcq	5	n/a	—	0.17	1.000	🏆 5/5 perfect

Key Findings

Finding 1: Perfect Quantization Parity — Zero Degradation Across All Families

This is the only model in the PBH Applied Systems evaluated series where quantization produces zero measurable behavioral change. Every pass rate, every bucket score, every signal rate is identical between F16 and Q4_K_M:

Family	F16 Pass Rate	Q4_K_M Pass Rate	Degradation
json_multistep	0.800	0.800	None
stateful_followup	1.000	1.000	None
toolcall_only	0.000	0.000	None
mixed_brief_json	1.000	1.000	None
toolcall	1.000	1.000	None
fuzz bucket	10.000	10.000	None
MCQ bucket	1.000	1.000	None

What this means for deployment: The Q4_K_M variant is a fully faithful quantization for behavioral purposes. There is no capability tradeoff at this precision level for this model — only a hardware and speed benefit (21.1× faster, ~3× less VRAM).

Finding 2: toolcall — Bucket=11 at Both Precision Levels

toolcall achieves the maximum possible bucket score (11.000) at both runners. This means the tool call JSON is correctly formed, schema-valid, and the final answer is correct and cleanly extracted.

Case	F16 Raw	Q4_K_M Raw	Expected
tool_01	`{...add(2,3)...} 5`	`{...add(2,3)...} 5`	`5` ✅
tool_02	`{...add(10,-4)...} 6`	`{...add(10,-4)...} 6`	`6` ✅

No EOS token contamination. No role token prefix. No final_mismatch. This is the only model in the evaluated series where toolcall achieves a clean pass with correct final answers at both precision levels. Every other model in the series has either final_mismatch from EOS contamination (Qwen Q4_K_M series), role-token contamination (Qwen F16 series), or no answer emitted (Mistral series).

Finding 3: MCQ — Perfect 5/5 at Both Precision Levels

Case	F16	Q4_K_M
mcq_01	✅ B	✅ B
mcq_02	✅ B	✅ B
mcq_03	✅ C	✅ C
mcq_04	✅ B	✅ B
mcq_05	✅ B	✅ B

Perfect MCQ performance at both precision levels. No A-bias, no empty output, no invalid choices. The same MCQ fixture that causes failures across other models in the series (including mcq_02 which exhibits systematic A-bias in smaller models) is answered correctly here.

Finding 4: json_multistep — Single Precision-Invariant Failure

Only ms_easy_02 fails, identically on both runners, with oracle_equiv_ok=0, checks_consistent_ok=1. This is the same fixture that also fails for Qwen2.5-7B at both precision levels — suggesting a model-agnostic characteristic of this specific test case rather than a 14B capability gap.

Case	F16	Q4_K_M	Secs (F16)	Secs (Q4)
ms_easy_01	✅	✅	108.93	5.63
ms_easy_02	❌	❌	125.48	6.78
ms_med_01	✅	✅	149.31	8.05
ms_med_02	✅	✅	144.80	7.86
ms_hard_01	✅	✅	137.51	7.98

Finding 5: toolcall_only — Schema Vocabulary Differs by Runner

Both runners fail toolcall_only on args_ok, but with distinct wrong schemas:

Runner	toolonly_01 raw	toolonly_02 raw
F16	`{"tool": "add", "left": 5, "right": 10}`	`{"tool": "add", "left": 25, "right": 75}`
Q4_K_M	`{"tool": "add", "input": {"x": 5, "y": 10}}`	`{"tool": "add", "input": {"numbers": [25, 75]}}`

The F16 model uses "left"/"right" as argument keys. The Q4_K_M model uses a nested "input" object with varying key names. Both use "tool" instead of "tool_name" as the outer key. Tool name recognition is perfect (1.000) at both runners — the model identifies "add" correctly. The arg schema vocabulary is the failure point in both cases. Providing the exact expected schema in the system prompt would resolve this.

Finding 6: F16 json_01 — 376-Second Outlier

json_01 at F16 takes 376.99 seconds — nearly 6× longer than the other three json cases (63–70s). This is the most extreme single-case timing outlier in the evaluated series. The output is correct (bucket=10) and the model produces valid JSON placement decisions. The outlier reflects the 1M context window's capacity for extensive generation on certain inputs — the model appears to generate far more internal content before settling on the brief JSON answer. The Q4_K_M runner takes 3.38s on the same case.

Signal-Level Diagnostics (Q4_K_M = F16 — Identical)

json_multistep

Signal	Rate	Notes
schema_ok	1.000	Perfect at both
checks_consistent_ok	1.000	Perfect at both
stop_semantics_ok	1.000	Perfect at both
oracle_equiv_ok	0.800	ms_easy_02 only — precision-invariant

stateful_followup

Signal	Rate
turn1_parse_ok	1.000
turn2_parse_ok	1.000
turn1_exact_match	1.000
turn2_exact_match	1.000

toolcall_only

Signal	Rate
tool_name_ok	1.000
args_ok	0.000

mixed_brief_json

Signal	Rate
answer_line_ok	1.000
json_parse_ok	1.000
schema_ok	1.000

Recommended Use Cases

✅ Deploy with Confidence (Q4_K_M)

Full-document and long-context processing — The defining deployment advantage. 1M token context enables entire codebases, contracts, books, and conversation histories in a single context window. No other model in the evaluated series approaches this capability.
Stateful multi-turn agents — Perfect 1.000 at Q4_K_M in 1.09 sec/case. Clean JSON state output.
Multi-step planning with external validation — 0.800 pass rate with perfect internal consistency. Use with oracle validation for production reliability.
Hybrid brief + JSON outputs — mixed_brief_json 1.000 at 1.01 sec/case.
Tool-calling with response scaffolding and final answer — toolcall at bucket=11 — tool dispatch valid, final answer correct, no cleanup required. The cleanest toolcall result in the series.
MCQ and single-choice extraction — Perfect 5/5. No A-bias, no empty output.
Structured JSON outputs (single-step) — json and fuzz both bucket=10.000.

⚠️ Use with Guardrails (Q4_K_M)

Easy-difficulty multi-step planning — ms_easy_02 fails at both precision levels. Add oracle validation.
Bare tool-call dispatch — toolcall_only fails on args schema vocabulary. Provide exact key names in system prompt.
Long-context inference on constrained hardware — At 128K+ token context, VRAM requirements scale significantly beyond the ~12 GB baseline. Plan for multi-GPU or CPU offload configurations.

Context Window vs. VRAM Guide (Q4_K_M)

Context Length	Approx. KV Cache	Total VRAM Needed	Recommended Hardware
8K tokens	~0.5 GB	~12 GB	T4 16 GB · RTX 3080
32K tokens	~2 GB	~14 GB	T4 16 GB · A10G
64K tokens	~4 GB	~16 GB	A10G 24 GB · RTX 4090
128K tokens	~8 GB	~20 GB	A10G 24 GB · RTX 4090
256K tokens	~16 GB	~28 GB	A100 40 GB · 2× A10G
512K tokens	~32 GB	~44 GB	A100 80 GB · multi-GPU
1M tokens	~64 GB	~76 GB	Multi-GPU · CPU offload

Set n_ctx to your actual working context length, not the model maximum.

Hardware Requirements

Configuration	VRAM Required	Notes
Q4_K_M · 8K context	~12 GB	T4 16 GB · RTX 3080
Q4_K_M · 128K context	~20 GB	A10G 24 GB · RTX 4090
Q4_K_M · 1M context	~76 GB	Multi-GPU / CPU offload
F16 (companion repo) · 8K context	~32 GB	A100 40 GB · RTX 4090

Usage

Installation

pip install llama-cpp-python huggingface_hub

For GPU acceleration (CUDA):

CMAKE_ARGS="-DGGML_CUDA=on" pip install llama-cpp-python --force-reinstall --no-cache-dir

Python — llama-cpp-python

python

from huggingface_hub import hf_hub_download
from llama_cpp import Llama

model_path = hf_hub_download(
    repo_id="pbhappliedsystems/qwen-2.5-14B-instruct-1m-gguf-Q4-K-M",
    filename="qwen-2.5-14B-instruct-1m-gguf-Q4-K-M.gguf"
)

llm = Llama(
    model_path=model_path,
    n_ctx=32768,      # Set to actual working context; supports up to 1M
    n_gpu_layers=-1,
    verbose=False,
)

response = llm.create_chat_completion(
    messages=[
        {
            "role": "system",
            "content": "You are a precise assistant. Follow instructions exactly and return structured outputs when requested."
        },
        {
            "role": "user",
            "content": "Analyze the following document and return a JSON object with keys: summary, key_entities, risk_level, action_items."
        }
    ],
    temperature=0.7,
    max_tokens=1024,
)

print(response["choices"][0]["message"]["content"])

For long-document processing (leveraging the 1M context window):

python

# Process a large document — adjust n_ctx to actual document token count
with open("large_document.txt", "r") as f:
    document = f.read()

# Estimate token count (~4 chars per token)
estimated_tokens = len(document) // 4
context_size = min(max(estimated_tokens + 2048, 8192), 1048576)

llm_long = Llama(
    model_path=model_path,
    n_ctx=context_size,
    n_gpu_layers=-1,
    verbose=True,  # Monitor memory during large context loads
)

response = llm_long.create_chat_completion(
    messages=[
        {"role": "system", "content": "You are a document analysis expert."},
        {"role": "user", "content": f"Summarize the following document and extract all action items:\n\n{document}"}
    ],
    temperature=0.7,
    max_tokens=2048,
)
print(response["choices"][0]["message"]["content"])

For tool-calling (clean results — no EOS stripping required at Q4_K_M):

python

# quant_eval v7.21: toolcall bucket=11 at both runners — clean final answer, no cleanup needed
response = llm.create_chat_completion(
    messages=[
        {
            "role": "system",
            "content": (
                "You are a tool-calling assistant. Output the tool call as JSON, "
                'then on the next line output only the numeric result.\n'
                'Tool call format: {"tool_name": "<n>", "args": {"a": <n>, "b": <n>}}'
            )
        },
        {"role": "user", "content": "Use the add tool to compute 10 minus 4."}
    ],
    temperature=0.7,
    max_tokens=128,
)
print(response["choices"][0]["message"]["content"])
# No stripping required — output is clean

For bare tool-call dispatch with explicit schema enforcement:

python

import json, re

def call_tool_bare(prompt: str, retries: int = 3) -> dict:
    """
    Bare tool dispatch with explicit schema.
    quant_eval v7.21: tool_name_ok=1.000, args_ok=0.000 — model uses 'left'/'right' or nested 'input'.
    Explicit schema in system prompt resolves the vocabulary issue.
    """
    for attempt in range(retries):
        response = llm.create_chat_completion(
            messages=[
                {
                    "role": "system",
                    "content": (
                        'Respond ONLY with a JSON object using EXACTLY these keys:\n'
                        '{"tool_name": "add", "args": {"a": <integer>, "b": <integer>}}\n'
                        'No other text, no markdown, no explanation.'
                    )
                },
                {"role": "user", "content": prompt}
            ],
            temperature=0.0,
            max_tokens=64,
        )
        raw = response["choices"][0]["message"]["content"].strip()
        try:
            parsed = json.loads(raw)
            assert "tool_name" in parsed and "args" in parsed
            assert "a" in parsed["args"] and "b" in parsed["args"]
            return parsed
        except (json.JSONDecodeError, AssertionError, KeyError):
            if attempt == retries - 1:
                raise ValueError(f"Tool call failed after {retries} attempts. Raw: {raw}")

result = call_tool_bare("Add 5 and 10.")

CLI — llama-cli

bash

llama-cli \
  --model qwen-2.5-14B-instruct-1m-gguf-Q4-K-M.gguf \
  --chat-template qwen2 \
  --system-prompt "You are a precise assistant. Follow instructions exactly." \
  --prompt "Analyze the following and return a JSON object with keys: summary, risk_level, action_items." \
  --n-predict 1024 \
  --ctx-size 32768 \
  --n-gpu-layers -1 \
  --temp 0.15

For server deployment:

bash

llama-server \
  --model qwen-2.5-14B-instruct-1m-gguf-Q4-K-M.gguf \
  --chat-template qwen2 \
  --ctx-size 32768 \
  --n-gpu-layers -1 \
  --port 8080 \
  --host 0.0.0.0

Query via the OpenAI-compatible API:

python

from openai import OpenAI

client = OpenAI(base_url="http://localhost:8080/v1", api_key="not-required")

response = client.chat.completions.create(
    model="qwen-2.5-14B-instruct-1m-gguf-Q4-K-M",
    messages=[{"role": "user", "content": "Your prompt here"}],
    temperature=0.7,
)
print(response.choices[0].message.content)

Evaluation Artifacts

The full per-case evaluation CSV (comparison_results_v7_21_Qwen2.5_14B_Instruct_1M_20260210_235131.csv) and rollup.json are published in this repository for independent verification.

Artifact Provenance

Artifact	Format	Size	SHA256
`qwen-2.5-14B-instruct-1m-gguf-Q4-K-M.gguf`	GGUF Q4_K_M	8.99 GB	`5ad529ff2b1b192f31c8a638fe8756a0c628904e2ded797c11f9194216976973`
F16 (companion repo)	GGUF F16	29.5 GB	`de08ea9c41234ef83b7aacf07f9ebc3cbaa20ca8aeb5f6417758a8798660aaa9`

Both artifacts were produced from Qwen/Qwen2.5-14B-Instruct-1M using a custom-built llama.cpp conversion and quantization pipeline developed by PBH Applied Systems.

Evaluation Methodology

quant_eval v7.21 — proprietary behavioral evaluation harness, PBH Applied Systems.

Fixture set: golden_oracle_fixtures_v7_21 (SHA256: 6d71a0b9147c079371b02a94f3c149eb78a6adc03dc16ff6833b964fbf4174f0)

Family	Description	Pass Signals
`fuzz`	Property-based regression; structured placement correctness	schema_ok, constraints_ok
`json`	Single-step structured JSON with constraint rules	schema_ok, constraints_ok
`json_multistep`	Multi-step planning with self-check and oracle verification	schema_ok, checks_consistent_ok, stop_semantics_ok, oracle_equiv_ok
`mcq`	Multiple-choice extraction	choice_ok
`stateful_followup`	Two-turn state tracking; turn-2 correct given turn-1	turn1/2_parse_ok, turn1/2_exact_match
`mixed_brief_json`	Hybrid: natural language answer + valid JSON block	answer_line_ok, json_parse_ok, schema_ok
`toolcall`	Tool call embedded in response; parse + schema validation	stage1_tool_parse_ok, stage1_tool_schema_ok
`toolcall_only`	Bare schema-only tool call; strict tool name + args check	tool_name_ok, args_ok

Evaluation hardware: NVIDIA RTX 4090 · Evaluation date: February 10, 2026 · Seed: 42

🔬 About quant_eval & This Evaluation Series

quant_eval is a proprietary 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 → The demo runs production-style agent workflows powered by open-weight models selected through the quant_eval evaluation pipeline.

Need a deployment recommendation? Not sure which quantization level is right for your hardware, latency target, or agent type? → pbhappliedsystems.com

Evaluated and published by PBH Applied Systems, LLC · patrick@pbhappliedsystems.com

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.

Patrick Hill, M.S. — Founder · Data Scientist · AI/ML Engineer · Author of Applied Machine Learning: Concepts, Tools, and Case Studies (required reading, UAT CSC 373)

Core Service Areas: LLM Optimization & Deployment · AI Evaluation Frameworks · Agentic AI Infrastructure · Scalable AI Application Development · ML Pipeline Design & Analytics · Model & Agent Cataloging

📞 Work With PBH Applied Systems

Qwen2.5-14B-Instruct-1M Q4_K_M is the highest-scoring model in the evaluated series across two dimensions, the only model with zero quantization degradation across all behavioral families, and the only model with clean toolcall final answers at both precision levels. That combination — top scores, no quantization penalty, 1M context, and 21× speedup at Q4_K_M — is what systematic evaluation documents. It's not visible from a leaderboard score or a casual test.

👉 Book a Scoping Call · 👉 Request an Evaluation Report — from $2,500

Connect


🌐	pbhappliedsystems.com
📧	patrick@pbhappliedsystems.com
💼	LinkedIn
▶️	YouTube
📸	Instagram
👍	Facebook

License

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

The quant_eval evaluation methodology, fixture set, and scoring framework are proprietary to PBH Applied Systems, LLC and are not included in this repository.

GGUF conversion, quantization, and behavioral evaluation performed by PBH Applied Systems, LLC · quant_eval v7.21 · Run ID: 20260210_235131

pbhappliedsystems/qwen-2.5-14B-instruct-1m-gguf-Q4-K-M

Qwen2.5-14B-Instruct-1M · GGUF Q4_K_M

Try This Model in the Live AI Agent Demo

Model Description

Key Characteristics

PBH Applied Systems Evaluation — quant_eval v7.21

Aggregate Scores (Q4_K_M)

Per-Family Pass Rates

F16 Baseline (full_weight_transformers)

Q4_K_M (quantized_llama_cpp)

Key Findings

Finding 1: Perfect Quantization Parity — Zero Degradation Across All Families

Finding 2: toolcall — Bucket=11 at Both Precision Levels

Finding 3: MCQ — Perfect 5/5 at Both Precision Levels

Finding 4: json_multistep — Single Precision-Invariant Failure

Finding 5: toolcall_only — Schema Vocabulary Differs by Runner

Finding 6: F16 json_01 — 376-Second Outlier

Signal-Level Diagnostics (Q4_K_M = F16 — Identical)

json_multistep

stateful_followup

toolcall_only

mixed_brief_json

Recommended Use Cases

✅ Deploy with Confidence (Q4_K_M)

⚠️ Use with Guardrails (Q4_K_M)

Context Window vs. VRAM Guide (Q4_K_M)

Hardware Requirements

Usage

Installation

Python — llama-cpp-python

CLI — llama-cli

Evaluation Artifacts

Artifact Provenance

Evaluation Methodology

🔬 About quant_eval & This Evaluation Series

About PBH Applied Systems

📞 Work With PBH Applied Systems

Connect

License

pbhappliedsystems/
qwen-2.5-14B-instruct-1m-gguf-Q4-K-M

F16 Baseline (`full_weight_transformers`)

Q4_K_M (`quantized_llama_cpp`)