---

license: apache-2.0

base_model: mistralai/Ministral-3-14B-Instruct-2512-BF16

tags:

- quant_eval

- pbh-applied-systems

- gguf

- bf16

- full-precision

- mistral

- ministral

- instruct

- llama-cpp

- agentic

- tool-calling

- structured-output

---

Ministral-3-14B-Instruct-2512 · GGUF BF16

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 v7.22.22 — 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 full-precision BF16 baseline repository. The evaluated quantized variants are published separately: Q4_K_M, Q5_K_M, Q8_0. This card documents reference behavior. The quantized cards show how each compression level impacts behavioral fidelity.

---

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 · 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
• Document Intelligence — Long-context information extraction and Q&A
• Code & Automation — API scaffolding, data transformation, and task completion

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 demo shows what this model can do. The Arena shows how this model compares to others. If you're deciding between BF16 and a quantized build, or between this model and another in the series, run the same query in the Arena with both selected.

Evaluation-backed results: every score shown for this model comes from its sealed quant_eval run bundles; the full results page is pbhappliedsystems.com/models/ministral-3-14b-instruct-2512.

---

Model Description

This repository contains the full-precision BF16 GGUF of mistralai/Ministral-3-14B-Instruct-2512-BF16, a 14B parameter instruction-tuned model developed by Mistral AI.

The BF16 format preserves the upstream checkpoint's weights without quantization. In the PBH Applied Systems evaluation pipeline, this BF16 build is the reference baseline: it ran in every run, against the same fixture set as each quantized build, enabling direct measurement of quantization impact.

Key Characteristics

• Parameters: 14B
• Format: GGUF BF16 (full precision)
• File size: 27.02 GB
• SHA256: 2593b4d01e56a79db0b835172773f5a984aca2b1de16de124ebe4d6459d864b6
• Evaluation context: 8,192 tokens (llama.cpp configuration)
• Evaluation hardware: NVIDIA A100-40GB
• Inference speed (eval hardware): 39.05, 38.05, 37.30 tokens/sec across the 3 runs of the identical BF16 artifact (observed harness throughput)

---

PBH Applied Systems Evaluation — quant_eval v7.22.22

> Evaluation conducted by PBH Applied Systems, LLC using quant_eval v7.22.22

> Runs: Ministral_3_14B_Instruct_2512_20261002_051323 (Q4_K_M) · Ministral_3_14B_Instruct_2512_20261002_190056 (Q5_K_M) · Ministral_3_14B_Instruct_2512_20261002_211552 (Q8_0)

> Seed: 42 · Hardware: NVIDIA A100-40GB

Evaluation Data

The results on this card are summaries of the sealed quant_eval run bundles. They are not deposited as datasets.

• Full results page — every family, every precision, with intervals and run provenance: pbhappliedsystems.com/models/ministral-3-14b-instruct-2512
• Per-case results — every case, both lanes, every gating signal: available on request
• Golden Oracle Fixtures — the fixture set used here (SHA-256 0673b5e5c0fb1b7b…) is the one published in the quant_eval Public Corpus, D3, DOI 10.5281/zenodo.22010278
• Run bundles — Q4_K_M Ministral_3_14B_Instruct_2512_20261002_051323, checksums.json SHA-256 4591195080b74172…; Q5_K_M Ministral_3_14B_Instruct_2512_20261002_190056, checksums.json SHA-256 406e765e9b2ce230…; Q8_0 Ministral_3_14B_Instruct_2512_20261002_211552, checksums.json SHA-256 9cfc827095e43cd3…

Evaluation Contract: canonical_agentic_contract_v7.22.18:evaluation

---

Comparability & Fixture Methodology

Identical Fixtures Across All Three Runs

| Run ID | Quantized Variant | Harness | Evaluation Contract | Fixture File SHA256 |

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

| Ministral_3_14B_Instruct_2512_20261002_051323 | Q4_K_M | v7.22.22 | v7.22.18:evaluation | 0673b5e... |

| Ministral_3_14B_Instruct_2512_20261002_190056 | Q5_K_M | v7.22.22 | v7.22.18:evaluation | 0673b5e... |

| Ministral_3_14B_Instruct_2512_20261002_211552 | Q8_0 | v7.22.22 | v7.22.18:evaluation | 0673b5e... |

Baseline reproducibility: The BF16 baseline was executed in all 3 runs. All eight family pass rates are identical across the runs, and at case level 0 of 1,600 cases differ in any pairwise comparison. Every quantized delta is measured against the BF16 lane of its own run.

---

Per-Family Pass Rates — Full-Weight Baseline

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

| Family | Pass Rate | 95% CI | Q4_K_M vs BF16 | Q5_K_M vs BF16 | Q8_0 vs BF16 |

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

| stateful_followup | 1.000 | [0.981, 1.000] | ±0.000 | ±0.000 | ±0.000 |

| mcq | 1.000 | [0.981, 1.000] | ±0.000 | ±0.000 | ±0.000 |

| toolcall | 0.950 | [0.910, 0.973] | ±0.000 | ±0.000 | ±0.000 |

| toolcall_only | 0.900 | [0.851, 0.934] | −0.010 | ±0.000 | +0.005 |

| fuzz | 0.600 | [0.531, 0.665] | ±0.000 | −0.025 | −0.030 |

| mixed_brief_json | 0.590 | [0.521, 0.656] | −0.075 ⚠️ | −0.030 | +0.030 |

| json_multistep | 0.505 | [0.436, 0.574] | ±0.000 | ±0.000 | ±0.000 |

| json | 0.460 | [0.392, 0.529] | +0.010 | +0.020 | +0.015 |

⚠️ = statistically significant loss at that precision (McNemar p < 0.05).

Source: sealed quant_eval run bundles; full tables at pbhappliedsystems.com/models/ministral-3-14b-instruct-2512

---

Key Findings

Finding 1: stateful_followup, mcq — at ceiling in BF16

| Family | Pass Rate | 95% CI |

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

| stateful_followup | 1.000 | [0.981, 1.000] |

| mcq | 1.000 | [0.981, 1.000] |

Context: every case passed at full precision. A quantized build can only hold or lose here, so these families show quantization loss most directly.

Finding 2: json and json_multistep — lowest pass rates at BF16

| Family | Pass Rate | 95% CI |

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

| json | 0.460 | [0.392, 0.529] |

| json_multistep | 0.505 | [0.436, 0.574] |

Context: these rates are measured at full precision, so they are characteristics of the model, not quantization artifacts. Agents that depend on these capabilities need scaffolding at any precision: json, json_multistep.

Finding 3: Quantization impact by precision

| Variant | Significant changes | Families affected |

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

| Q4_K_M | 1 of 8 | mixed_brief_json (−0.075) |

| Q5_K_M | 0 of 8 | — |

| Q8_0 | 0 of 8 | — |

---

Quantization Variants Comparison

See how this model performs across all three quantized variants. All three were evaluated on identical fixtures against the same BF16 baseline.

| Variant | File Size | Speed | Stateful | Toolcall-only | Hybrid | Significant changes | Best For |

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

| BF16 | 27.02 GB | 1.0× | 1.000 | 0.900 | 0.590 | — | Reference baseline; maximum fidelity |

| Q4_K_M | 8.24 GB | 1.49× | 1.000 | 0.890 | 0.515 | 1 of 8 | Smallest footprint; check mixed_brief_json |

| Q5_K_M | 9.62 GB | 1.35× | 1.000 | 0.900 | 0.560 | 0 of 8 | No significant change from BF16 |

| Q8_0 | 14.36 GB | 1.39× | 1.000 | 0.905 | 0.620 | 0 of 8 | No significant change from BF16 |

How to read this table: Speed is the observed evaluation wall-time ratio (BF16 ÷ quantized) on matched A100-40GB hardware — not a controlled throughput benchmark. Stateful/Toolcall-only/Hybrid are pass rates.

See the full cards:

• Q4_K_M
• Q5_K_M
• Q8_0

---

Artifact Provenance

| Artifact | Format | Size | SHA256 |

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

| ministral-3-14b-instruct-2512-gguf-F16.gguf | GGUF BF16 | 27.02 GB | 2593b4d01e56a79db0b835172773f5a984aca2b1de16de124ebe4d6459d864b6 |

Produced from mistralai/Ministral-3-14B-Instruct-2512-BF16 at revision 3cea74c1ebaf5ce5f5a2553de470e2ceab825142 (exact) using a custom-built llama.cpp conversion and quantization pipeline developed by PBH Applied Systems.

Artifact verification: the SHA-256 is recorded in the run manifest and checked by the evaluation runtime when the file is loaded. Download the GGUF and verify:

sha256sum ministral-3-14b-instruct-2512-gguf-F16.gguf
# Should output: 2593b4d01e56a79db0b835172773f5a984aca2b1de16de124ebe4d6459d864b6

---

Evaluation Methodology

quant_eval v7.22.22 is a behavioral evaluation harness developed by PBH Applied Systems. Each run evaluates the full-precision (BF16) model and one quantized variant against the same fixture set, enabling direct measurement of behavioral degradation. Method: Whitepaper 1 (the harness) and Whitepaper 2 (the public corpus that validates it).

Fixture set: v7.22.22 evaluation split

• SHA256: 0673b5e5c0fb1b7ba4bdc8c9aa985104260db183453c5708aff14c87823edb97
• Total unique cases: 1,600 (200 per family × 8 families)

| Family | Description | Pass Signals |

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

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

| 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_parse_ok, turn2_parse_ok, turn1_exact_match, turn2_exact_match |

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

| mcq | Multiple-choice extraction | semantic_choice_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 |

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

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

Evaluation hardware: NVIDIA A100-40GB

Runner: full_weight_modal_llama_cpp (llama.cpp via Python binding on Modal)

Decoding: temperature 0.15, seed 42, top_p 0.95, top_k 40

Context: 8,192-token context window (llama.cpp configuration)

Reproducibility: sealed run bundles (checksums above); per-case results on request

---

Deployment Recommendations

✅ Deploy BF16 when:

• Maximum fidelity matters more than cost. This is the reference every quantized build is measured against.
• You need to reproduce or audit this evaluation. The BF16 lane gave identical pass counts in every run.
• You can afford 27.02 GB of weights on A100-40GB-class hardware or larger.

⚠️ Avoid BF16 if:

• A quantized build fits your workload. Q5_K_M (9.62 GB) and Q8_0 (14.36 GB) showed no statistically significant change in any family.

---

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:

• Full-Weight vs Quantized Model Evaluation Report — A written behavioral audit: per-family pass rates, full-weight vs. Q4_K_M, Q5_K_M and Q8_0 delta analysis, deployment recommendation, and the lowest precision at which each capability holds. Evidence-standard numbers only. Typically $2,500–$5,000. Read a complete sample report (this model, as a client receives it).
• Per-Case Results Request — the per-case results behind this card, sent in writing: request them.
• 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 — mistralai/Ministral-3-14B-Instruct-2512-BF16

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 conversion and behavioral evaluation performed by PBH Applied Systems, LLC · quant_eval v7.22.22