Ministral-3-14B-Reasoning-2512 · GGUF F16

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.

📌 This is the full-precision F16 baseline repository. The evaluated Q4_K_M deployment variant is published at pbhappliedsystems/ministral-3-14b-reasoning-2512-gguf-Q4-K-M. That card documents the complete F16 vs. Q4_K_M comparison — including the 55.7× inference speedup and the hard planning case capability regression — in full detail.

🧠 This is a reasoning-architecture model. Ministral-3-14B-Reasoning-2512 generates extended internal chain-of-thought traces before producing its final answer. At full F16 precision, this reasoning capability is fully expressed — and measurable. The evaluation data below documents exactly where it helps, where it creates extraction friction, and why it matters for deployment decisions.

---

Try the Live AI Agent Demo

Launch the PBH Applied Systems AI Agent Demo →

This model is part of the PBH Applied Systems evaluated model series that supports the live AI Agent Demo. The demo lets visitors interact with production-style agent workflows powered by open-weight language models evaluated through PBH Applied Systems' quant_eval framework.

The F16 model serves a different role than the Q4_K_M deployment variant. F16 is the full-precision baseline used to measure what the model can do before quantization. quant_eval then compares the quantized model against this baseline to identify which capabilities are preserved, which degrade, and which tasks require guardrails or a higher-precision deployment.

This comparison is central to the demo. It helps determine which model belongs in which agent role:

• Reasoning models are selected for planning, analysis, and auditable decision workflows.
• Document models are selected for long-context extraction, summarization, and structured Q&A.
• Code models are selected for task completion, structured output, API scaffolding, and automation workflows.
• Quantized variants are selected when they preserve enough behavior to reduce cost, latency, and GPU requirements.
• F16 variants remain important when maximum fidelity, cleaner tool execution, or reduced quantization risk matters more than speed or cost.

The live demo shows the deployment side of that process. The F16 card documents the reference behavior. The Q4_K_M card shows what changes after compression. Together, they explain how PBH Applied Systems uses quant_eval to choose the correct LLM for the correct agent type instead of guessing from model size or leaderboard reputation.

---

Model Description

This repository contains the full-precision F16 GGUF of mistralai/Ministral-3-14B-Reasoning-2512, a 14-billion parameter reasoning-tuned model from Mistral AI (December 2025 release).

The F16 format preserves all original float16 weights without quantization, enabling the model's full chain-of-thought capability. In the PBH Applied Systems evaluation pipeline, this F16 run (20260209_212510) served as the baseline cache generation pass — producing the full_weight_cache.json used as the reference anchor for the subsequent Q4_K_M comparison run (20260209_233252). This two-pass architecture ensures that F16 and Q4_K_M results are measured against the identical fixture set under controlled, reproducible conditions.

For most production deployments, the Q4_K_M variant is the appropriate choice due to its dramatically lower VRAM requirement and 55.7× faster inference. The F16 is the right choice when the full chain-of-thought is required — for hard planning tasks, deliberative reasoning workflows, or any use case where the thinking trace itself is part of the deliverable.

Key Characteristics

• Parameters: 14B
• Architecture: Reasoning (extended chain-of-thought at full precision)
• Format: GGUF F16 (full precision)
• File size: 27.0 GB
• SHA256: 7645d01deed3415326c7c2bf8b58280e234021a91b4b3ade52b4735c976ad221
• Minimum VRAM (GPU inference): ~30 GB
• Recommended GPU tier: A100 40 GB · RTX 4090 (24 GB, with offload) · 2× A10G
• Context window: 32,768 tokens (per base model specification)
• Inference speed (eval hardware): avg 65.67 sec/case on RTX 4090

Why 65.67 sec/case? This is the cost of full reasoning at F16. The model generates extended chain-of-thought traces before each response. The json family averaged 105.82 sec/case; the fuzz suite peaked at 176.47 sec on a single case. This is not a performance problem — it is the reasoning capability operating as designed at full precision. For comparison, the Q4_K_M variant averages 1.18 sec/case with the chain-of-thought substantially compressed.

---

PBH Applied Systems Evaluation — quant_eval v7.21

Evaluation conducted by PBH Applied Systems, LLC using quant_eval v7.21 Run ID: 20260209_212510 · Fixtures: golden_oracle_fixtures_v7_21 (SHA256: 6d71a0b9147c...) · Seed: 42 Hardware: NVIDIA RTX 4090 · Runner: full_weight_transformers (F16 only) · Total rows: 42

Note on aggregate scores: The normalized aggregate dimensions (task completion, reasoning, coherence, instruction following) are computed from the combined F16 + Q4_K_M comparison run and are reported on the Q4_K_M card. This F16 card reports per-family pass rates from the full_weight_transformers runner — the authoritative signal for understanding full-precision capability.

Per-Family Pass Rates — F16 (full_weight_transformers)

json_multistep — Case-Level Breakdown

The F16 passes the hard case. ms_hard_01 — the most demanding planning case in the fixture set — is solved correctly by the F16 model in 110.73 seconds of deliberation. The Q4_K_M variant fails this same case (checks_consistent_ok=0, oracle_equiv_ok=0) in 3.39 seconds. This is the clearest empirical argument for deploying F16 when planning correctness on harder cases is a requirement.

The easy case failures (ms_easy_01, ms_easy_02) are a pattern specific to the Reasoning architecture: the model's extended deliberation on simple cases sometimes leads it to over-engineer the solution, producing an oracle_equiv_ok failure where the Instruct variant succeeds without the reasoning overhead. The chain-of-thought is a double-edged signal — it helps on hard cases and occasionally over-complicates easy ones.

⚠️ toolcall_only — Architectural Limitation (Both Runners)

toolcall_only fails at 0.000 on the F16 runner, and also at 0.000 on the Q4_K_M runner. This is not a quantization finding — it is an architectural characteristic of the Reasoning model that applies at both precision levels.

Case-level detail:

• toolonly_01 (15.47s): tool_name_ok=1, args_ok=0, schema_ok=0 — tool is identified but args extraction fails amid reasoning output
• toolonly_02 (13.34s): tool_name_ok=0, args_ok=0, schema_ok=0 — the reasoning trace fully interferes with tool name extraction

The chain-of-thought output structure is incompatible with bare schema-only tool dispatch at full precision. The toolcall family (tool call embedded within a broader response) passes at 1.000 — the reasoning traces are absorbed naturally into the response structure when there is surrounding prose scaffolding.

Practical implication: Use response-scaffolded tool calling at both precision levels. See the Q4_K_M card for the call_tool_scaffolded() implementation pattern.

MCQ — Backtick Wrapping Artifact

mcq_02 fails with detail=invalid_choice raw='```' at 3.97 seconds. The F16 model emits a markdown code fence delimiter as its extracted response — the reasoning trace wraps the answer in a code block before the choice letter, and the extractor captures the fence instead of the letter.

All other MCQ cases pass correctly at 3.97–4.12 sec each. This is a single-case formatting artifact, not a general MCQ capability failure. A post-processing step that strips markdown fences before choice extraction would resolve it.

Signal-Level Diagnostics (F16)

json_multistep

checks_consistent_ok = 1.000 at F16 is a meaningful signal. The Q4_K_M variant drops this to 0.800. At full precision, the model's internal reasoning maintains self-consistent intermediate check states across all five cases — it is only oracle equivalence (final placement correctness) that fails on the two easy cases.

stateful_followup

toolcall_only

mixed_brief_json

---

When to Deploy F16 vs. Q4_K_M

---

Recommended Use Cases — F16

✅ Deploy with Confidence

• Hard multi-step planning — The only variant in this pair that passes ms_hard_01. Use F16 when planning horizon difficulty is unpredictable or when hard cases must be handled correctly.
• Deliberative reasoning workflows — Use cases where the chain-of-thought trace itself is the deliverable: audit trails, decision documentation, explainable AI pipelines.
• Stateful multi-turn agents — Perfect two-turn state retention (1.000). The reasoning architecture maintains context across turns reliably at full precision.
• Hybrid brief + JSON outputs — mixed_brief_json passes at 1.000. The reasoning traces do not interfere with mixed-format output at full precision.
• Scaffolded tool-calling — toolcall passes at 1.000. Embed tool dispatch within a broader response structure.
• Structured single-step JSON — json and fuzz both achieve bucket_score avg of 10.000, with perfect constraint adherence despite the 105-second deliberation cost.

⚠️ Use with Extraction Normalization

• MCQ and single-choice outputs — Add a post-processing step to strip markdown fences before choice extraction. One of five cases fails due to backtick wrapping; the underlying answer is present but embedded in a code block.
• Multi-step planning at easy difficulty — The easy cases (ms_easy_01, ms_easy_02) fail at F16 due to reasoning over-engineering. Use with an external validator or prefer Q4_K_M for easy-difficulty tasks.

❌ Not Recommended

• Bare tool-call dispatch (schema-only output) — toolcall_only fails at 0.000. This is an architectural characteristic, not a precision issue. Use scaffolded tool calling instead.
• High-throughput pipelines — At 65.67 sec/case average and up to 176 seconds for fuzz cases, F16 is not suitable for latency-sensitive or batch-at-scale workloads.

---

Hardware Requirements

---

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

from huggingface_hub import hf_hub_download
from llama_cpp import Llama

# Note: 27 GB download — ensure sufficient disk space and ~30 GB VRAM
model_path = hf_hub_download(
    repo_id="pbhappliedsystems/ministral-3-14b-reasoning-2512-gguf-F16",
    filename="ministral-3-14b-reasoning-2512-gguf-F16.gguf"
)

llm = Llama(
    model_path=model_path,
    n_ctx=16384,      # Increase for long reasoning traces; up to 32768 per model spec
    n_gpu_layers=-1,  # -1 offloads all layers to GPU; reduce if VRAM < 30 GB
    verbose=False,
)

response = llm.create_chat_completion(
    messages=[
        {
            "role": "system",
            "content": (
                "You are a precise reasoning assistant. Think through each problem step-by-step "
                "before giving your final answer."
            )
        },
        {
            "role": "user",
            "content": "Analyze the following contract for logical inconsistencies and conflicting obligations. Return a structured JSON summary with keys: inconsistencies, obligations, risk_level, recommendation."
        }
    ],
    temperature=0.7,
    max_tokens=4096,  # Reasoning traces at F16 can be long — allocate generously
)

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

For partial GPU offload when VRAM is between 16–24 GB:

llm = Llama(
    model_path=model_path,
    n_ctx=8192,
    n_gpu_layers=20,   # Tune based on VRAM; remainder offloads to CPU
    verbose=True,      # Enable to monitor layer offload and memory usage
)

For MCQ with extraction normalization (addresses the backtick wrapping artifact noted above):

import re
from huggingface_hub import hf_hub_download
from llama_cpp import Llama

model_path = hf_hub_download(
    repo_id="pbhappliedsystems/ministral-3-14b-reasoning-2512-gguf-F16",
    filename="ministral-3-14b-reasoning-2512-gguf-F16.gguf"
)

llm = Llama(model_path=model_path, n_ctx=4096, n_gpu_layers=-1, verbose=False)

def extract_mcq_choice(raw: str, valid_choices: list = ["A", "B", "C", "D"]) -> str:
    """
    Extract MCQ choice from reasoning model output.
    Strips markdown fences and reasoning traces before extraction.
    Addresses quant_eval v7.21 finding: mcq_02 fails with invalid_choice raw='```'
    """
    # Strip markdown fences
    clean = re.sub(r'```[^\n]*\n?', '', raw).strip()
    # Find last occurrence of a valid choice letter (reasoning may mention choices mid-trace)
    for char in reversed(clean.split()):
        candidate = char.strip('.,;:()\'"').upper()
        if candidate in valid_choices:
            return candidate
    raise ValueError(f"No valid choice found in: {raw[:200]}")

response = llm.create_chat_completion(
    messages=[
        {"role": "system", "content": "Answer with only the letter of the correct choice."},
        {"role": "user", "content": "Which of the following is a primary color? A) Green B) Orange C) Blue D) Purple"}
    ],
    temperature=0.0,
    max_tokens=512,
)
raw = response["choices"][0]["message"]["content"]
choice = extract_mcq_choice(raw)
print(f"Extracted choice: {choice}")

CLI — llama-cli

# Deliberative reasoning prompt (expect 30–120 second response time at full precision)
llama-cli \
  --model ministral-3-14b-reasoning-2512-gguf-F16.gguf \
  --chat-template mistral \
  --system-prompt "You are a precise reasoning assistant. Think step-by-step before responding." \
  --prompt "Analyze the following multi-step decision problem and return a JSON plan with checks at each step." \
  --n-predict 4096 \
  --ctx-size 16384 \
  --n-gpu-layers -1 \
  --temp 0.15

For server deployment:

llama-server \
  --model ministral-3-14b-reasoning-2512-gguf-F16.gguf \
  --chat-template mistral \
  --ctx-size 16384 \
  --n-gpu-layers -1 \
  --port 8080 \
  --host 0.0.0.0

Query via the OpenAI-compatible API:

from openai import OpenAI

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

response = client.chat.completions.create(
    model="ministral-3-14b-reasoning-2512-gguf-F16",
    messages=[{"role": "user", "content": "Your prompt here"}],
    temperature=0.7,
    timeout=300,  # F16 reasoning responses may take 60–180 seconds
)
print(response.choices[0].message.content)

---

Artifact Provenance

The F16 GGUF was converted from the mistralai/Ministral-3-14B-Reasoning-2512 HuggingFace snapshot using a custom-built llama.cpp conversion pipeline developed by PBH Applied Systems, without modification to model weights prior to conversion.

Two-pass evaluation architecture: The F16 evaluation run (20260209_212510) operated in cache-generation mode (skip_quant=true), producing the full_weight_cache.json that served as the reference baseline for the subsequent Q4_K_M comparison run (20260209_233252). This architecture ensures that F16 and Q4_K_M results are measured against the identical fixture set under reproducible, controlled conditions, and that the F16 results cached during the first pass are the exact same responses compared against in the second pass.

---

Evaluation Methodology

quant_eval v7.21 is a proprietary behavioral evaluation harness developed by PBH Applied Systems. The two-run architecture evaluates the full-precision (F16) model first, caches its results, then evaluates the quantized variant against the same fixture set — enabling exact apples-to-apples comparison.

Fixture set: golden_oracle_fixtures_v7_21 (SHA256: 6d71a0b9147c079371b02a94f3c149eb78a6adc03dc16ff6833b964fbf4174f0)

Evaluation hardware: NVIDIA RTX 4090 (24 GB VRAM) F16 evaluation date: February 9, 2026 quant_eval 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. The organization operates with a strong emphasis on 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

PBH Applied Systems was founded by Patrick Hill, a Data Scientist and AI/ML Engineer 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:

• Languages & Data: Python, SQL, Linux, Pandas, NumPy, scikit-learn
• ML & Modeling: Supervised and unsupervised learning, neural networks, NLP, transformers, regression, classification, forecasting, and feature engineering
• AI/ML Frameworks: PyTorch, TensorFlow/Keras, HuggingFace Transformers, GGUF, llama.cpp, BitsAndBytes, PEFT, QLoRA
• Deployment & MLOps: Flask APIs, Docker, CI/CD pipelines, REST endpoints, streaming inference, version control
• Data Platforms: Jupyter, Databricks, Power BI, Matplotlib
• Quantization: GGUF conversion, Q4_K_M / Q5_K_M / Q8_0 strategies, adapter-per-model evaluation architecture

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

1. LLM Optimization & Deployment — End-to-end GGUF conversion and quantization with custom llama.cpp pipelines and adapter-per-model architecture.

2. AI Evaluation Frameworks — Proprietary behavioral evaluation via quant_eval: per-family pass rates, F16 vs. quantized delta analysis, failure cluster diagnostics, deployment recommendations.

3. Agentic AI Infrastructure — LlamaIndex ReAct agents, Flask orchestration, serverless GPU inference, full pipeline from model selection to production serving.

4. Scalable AI Application Development — Multimodal applications (quantized LLMs + Whisper + BLIP), Dockerized Flask APIs, advanced time-series forecasting with custom attention mechanisms, Bayesian hyperparameter optimization, and FinBERT sentiment fusion.

5. ML Pipeline Design & Analytics — Feature engineering, forward-chaining cross-validation, KPI dashboards, analytical governance at scale.

6. Model & Agent Cataloging — Structured catalog publishing with reproducible artifacts and clear performance tradeoff documentation.

---

📞 Work With PBH Applied Systems

This F16 card documents the full reasoning capability of Ministral-3-14B-Reasoning-2512 at maximum precision. The Q4_K_M companion card shows what changes when you quantize — including which capabilities survive the 55.7× speedup and which do not. The hard planning case that takes 110 seconds to solve correctly at F16 fails in 3 seconds at Q4_K_M. That difference is only visible when you evaluate both.

👉 Book a Scoping Call — Discuss your reasoning model deployment strategy, quantization tradeoffs, or agentic architecture directly with Patrick.

👉 Request an Evaluation Report — A full quant_eval behavioral audit: per-family pass rates, F16 vs. quantized delta analysis, failure cluster diagnostics, and a deployment recommendation. Engagements from $2,500.

Connect

---

License

This GGUF repository inherits the license of the base model: Apache 2.0 — mistralai/Ministral-3-14B-Reasoning-2512

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 and behavioral evaluation performed by PBH Applied Systems, LLC · quant_eval v7.21 · F16 Run ID: 20260209_212510