Skip to content
CrucibleMark

Ministral 3 14B (Unsloth)

Ministral 3 14B is the top model of the Ministral 3 family for local assistant and analysis workloads. 13.9B dense parameters, 256,000 tokens of context, multimodal input for text and image, native function calling and JSON output. Licensed under Apache 2.0 and fully operable locally as an Unsloth GGUF variant — the most capable Ministral to date without any cloud dependency.

Mistral AI Version 3 Commercial use permitted Dense 13.9 B 256 K Context 07/2025 locally tested

  • Open Weights
  • Desktop
  • llama.cpp
  • Text
  • Vision
  • Instruction-Tuned
  • Batch

Sovereign Risk: LOW Mistral AI releases Ministral 3 under Apache 2.0, and Unsloth provides local GGUF packaging for offline inference. Since the upstream weights are open and the benchmark can run entirely locally, the operational provenance risk is low.

Tool-use profile: 6 assets in detail

Comparison of asset performance (P1/P2/Combined) against the fleet average

Asset performance (radar)

Score Breakdown vs. Fleet Average

Tool-use details

Asset performance, reliability, and runtime profile

Name P1 P2 Combined
EU License Research 75 40 54
HTTP Fetch & Extract 80 35 57.5
Tool Failure Handling (404) 80 15 57.3
Web Search & Tool Selection 100 35 70.8
URL Construction & Fetch 80 60 70
Multilingual Search & Synthesis 100 15 57.5

CrucibleMark evaluates tool use across 6 independent tests. Click a test name for details.

Reliability

  • Tool Call Valid: No
  • Retry: Not required
  • Hallucination: Detected

Reliability measures how consistently a model actually executes tool calls: Tool Call Valid schema and format accepted, Retry Required successful only after retry, Hallucination Flag fabricated tools or parameters detected. All three green means production-ready.

Runtime profile

Call 1
4.91
First Request
MCP
1.75
Protocol Latency
Synthesis
63.29
Response Generation
Total
419.68
Sum of All Phases
Token
12144
Input + Output
Cost
$0
Cost per Run

The runtime profile shows the latency and cost metrics for the model run: Call 1 First Request, MCP Protocol Latency, Synthesis Response Generation, Total sum of all phases, supplemented by Token input and output and Cost cost per run.

Tool-use review

· Instruction-Tuned

Deployment Verdict

Conditional deploy: The model shows usable tool execution, but is not reliable enough for autonomous MCP pipelines due to detected hallucination and an invalid tool call.

Tool Execution Profile

Ministral 3 14B fundamentally understands when it needs to use a tool. In the Web Search & Tool Selection test, which checks whether web_search is chosen over fetch without prompting, it makes the tool selection confidently. This argues against a purely rigid pattern. In the URL Construction & Fetch test, which measures the derivation of a correct target URL from internal knowledge, it generally performs adequately, but not precisely enough for strictly deterministic flows. The P1 score is solid overall, but the finding “Tool-Call valid: false” is severe in production use. The model can make sensible tool decisions, but does not consistently keep the handoff to infrastructure protocol-clean.

Synthesis Fidelity

How well does it consolidate tool results? Poorly. P2 performance at 33.33 falls clearly behind execution. The pattern across assets is consistent: fetching and search work more reliably than the subsequent clean consolidation. Particularly critical is Multilingual Search & Synthesis, which tests cross-border research and German-language summarization, with very low synthesis quality. For architectures where the model only needs to pass on tool results or structure them briefly, this is still manageable. For analytical final responses, it is not.

Does it stay within the tool result or fall back on training data? In the Honeypot EU License Research test, which distinguishes current license restrictions from memorized knowledge, it at least stays within the correct behavioral space and does not hallucinate there. This is a positive trust signal for compliance-adjacent research. However, it is offset by the global hallucination finding. Once a model outputs fabricated facts as result content in a tool pipeline, it creates a security risk for the entire chain.

Error Resilience

Here the model fails in a production-critical way. In the Tool Failure Handling (404) test, which checks whether a failed retrieval is communicated openly, it produces fabricated page content despite the 404. This is not merely a quality weakness — it is a breach of trust. Transparent error communication would be acceptable. Hallucinated substitute content is not.

Sovereignty Profile

Locally deployable, Apache-2.0-licensed, and therefore attractive for sovereign deployments. Combined performance sits 6.58 points below the fleet average of 67.75. That is respectable for a Desktop model, but not strong enough to offset the reliability risks.

Conclusion & Recommendation

The model is suitable for locally operated, low-cost assistance pipelines with human oversight, clear guardrails, and tight scope limited to tool selection, research initiation, and structured preparatory work. It is not suitable for autonomous MCP orchestration, compliance outputs, error paths without supervision, or any pipeline in which tool results are directly processed as a trusted factual basis. Deploy only with a hard response validation layer and an explicit hallucination brake.

This analysis was generated automatically based on benchmark data. Model used: GPT-5.4 by OpenAI. The raw data and full methodology are documented in the GitHub project.