Mistral Medium 3.5

LLM Model Review

· Instruction-Tuned

With an overall score of 72.35%, Mistral Medium 3.5 demonstrates the rare combination of speed and usable breadth you’d expect from a commercial cloud model in Mistral AI’s API at this tier. The model is classified as a generalist, but clearly bears the hallmarks of its additional roles: Instruct in form, agentic in structure, multimodal in ambition. The Speed Profile Badge reads Real-Time Tool Expert, accompanied by 144.22 tokens per second. That’s not a fair-weather figure — it’s a clear signal: this model doesn’t want to deliberate like an essayist in productive tool operation; it wants to deliver like a well-oiled assistant. Sovereign Risk: LOW — Mistral AI is headquartered in France, processes data in the EU according to the Vendor Card, and is not subject to US CLOUD Act logic.

Header Grades: Stability and Reliability

Architecture and Character: What This Model Wants to Be

The editorial classification aligns surprisingly well with observed behavior. Mistral Medium 3.5 is a generalist — not a specialized tool with tunnel vision, but a model built for the full range of tasks. At the same time, it is classified as Server class. That means: no excuses from the lightweight division apply here. A dense architecture with 128 billion parameters and 128 billion active parameters doesn’t need to dominate everything, but it should hold its own seriously in almost every module. Mistral Medium 3.5 does exactly that. Just not always with the same degree of conviction.

The additional labels explain the tone of its responses. As an Instruct model, it follows instructions mostly cleanly and without rhetorical swings. As an agentic model, it often thinks in useful structures for tool and workflow contexts. This is especially visible where tasks decompose into steps, priorities, or production logic. The multimodal label should be properly bounded: this benchmark measures almost exclusively text performance. Anyone drawing a comprehensive verdict on image understanding or visual agent pipelines from these results would be claiming more than the data supports.

Performance and Price: Fast, but Not Cheap Enough to Forgive Mistakes

144.22 tokens per second is strong for a cloud model of this size. Combined with the Real-Time Tool Expert badge, a clear deployment profile emerges: Mistral Medium 3.5 is built for interactive work with tools, structured assistance, and rapid feedback loops. Not for contemplative long-haul tasks, but for the rhythm of everyday work.

The price is $1.5 per million input tokens and $7.5 per million output tokens. That’s not a bargain price. Given the speed and broad performance profile, it’s defensible — but not sensational. The benchmark cost of $0.3884 for the entire run stays within bounds, largely because the model works token-economically. No module exceeds the expected verbosity range. On the contrary: in almost every area, Mistral Medium 3.5 stays below the fleet median. It rarely writes too much. In a vendor cloud, that’s not just a matter of style — it’s simply a cost question.

Code Quality: Strong in Breadth, Slightly Weaker in Forensic Precision

With 79.6 points in the Code Quality Audit, Mistral Medium 3.5 performs visibly above average. The qualitative analysis confirms this impression. In a security analysis, the model identified 24 vulnerabilities — actually more than the reference standard. That’s a good sign initially: it sees a lot and isn’t intimidated by complex material.

The catch lies in the second step. Mistral Medium 3.5 is more of a thorough inventory-taker than an uncompromising incident analyst. It cleanly identifies critical issues such as SQL Injection, plaintext passwords, path traversal, IDOR, insecure cookies, and weak token generation. The formatting holds up too. The Markdown table was correct, the categories were usable, and the remediation suggestions were practical rather than decorative.

What was missing was the final degree of sharpness. The Judge rightly notes that the model lacked the concrete attack path — the chaining of multiple vulnerabilities into a realistic exploit scenario. That’s precisely where a good finding separates itself from a security-relevant prioritization. Even more notable is the miscalibration on Type Juggling. The model recognized the issue but rated it too low. For PHP-adjacent security work, that’s not a minor detail — it’s a warning sign: Mistral Medium 3.5 knows the map, but not every mine gets the right warning marker.

For code reviews, secure refactoring guidance, and structured vulnerability lists, this is strong. For high-stakes security audits without human review, it falls short. A model that spots security issues but doesn’t articulate their exploit chain clearly enough is useful. It’s just not a digital pentester yet.

CLI, Tool Use, and Agentic Practice: Where the Real Talent Shows

The Real-Time Tool Expert badge wasn’t assigned arbitrarily. In the CLI benchmark, Mistral Medium 3.5 scores 87.22 points; in Tool Execution, 90.0. This is the zone where its classification as an agentic model gains substance. The model decomposes tasks usably, works in a structured manner, and moves through tool contexts with a naturalness that many pure chat models lack.

But there is a flaw, and it weighs more heavily than a botched command. In one tool-use task, the model hallucinated content that did not originate from the retrieved tool result. The system therefore capped the P2 score via the hallucination cap. For content-critical tasks — research, audit reports, or fact-bound agent chains — this is a disqualifying signal. A model that calls a tool and then improvises freely afterward isn’t behaving like an assistant; it’s behaving like an intern with too much self-confidence.

This significantly relativizes the otherwise strong tool scores. In operational workflows with clear downstream verification, Mistral Medium 3.5 remains attractive. In pipelines where tool output is assumed to be true and passed on unverified, it needs to be kept on a shorter leash.

Reasoning and Logic: Right Answer, Unnecessarily Many Loops

The Logical Reasoning section ends with a score of 65.87 points. That’s not a collapse, but it’s no badge of honor for a Server-class model of this size. The qualitative probe is revealing: on a classic guard puzzle, Mistral Medium 3.5 found the correct solution, formulated the right meta-question, and reached the logical conclusion. Substantively, it passed.

The problem lies in the nature of its thinking. The documented reasoning was long, repetitive, and circled the same core multiple times. The Judge describes it as circular rumination. That’s apt. The model doesn’t think incorrectly — it often thinks like someone who draws the same sketch on the whiteboard five times over. This makes the solution feel more labored than it needs to be. The final explanations also remain terser and less didactically rich than they could be. Anyone deploying the model for clean, instructive derivations will get correct results — just not always elegant ones.

Metacognition Compliance (Reasoning): The model refuses to use the explicitly requested <thought> tags in 3/5 metacog tests, citing a consistent policy statement. The reasoning content itself is partially correct — the score deduction results from the format refusal, not from logical errors. For comparison: in the tag-free reasoning_5* tests, the model achieves an average score of approximately 65.87%, which is on par with other models. CrucibleMark deliberately evaluates native zero-shot instruction compliance as a real-world characteristic — this deduction is methodologically intentional.

This matters because it reveals the model’s character. Mistral Medium 3.5 doesn’t fail here primarily due to logic, but due to its willingness to follow a very specific formatting instruction without compromise. For regular users, this is less dramatic than for agent frameworks that expect precisely defined response formats.

Content Transformation: Functionally Strong, Not Top-Tier Emotionally

With 75.5 points, Content Transformation is one of the model’s stronger areas. The qualitative probe of a German-language video script rewrite shows a familiar pattern: Mistral Medium 3.5 completes the task fully and cleanly, but remains slightly too sober where more dramaturgical pressure would have helped.

The analysis of the source material was correct but brief. Missing elements such as a hook, spoken-word tone, production notes, and engagement components were named but not explained in depth. In the actual transformation, the model then delivered a functional script with timestamps, stage directions, a CTA, and even an Easter egg. That’s more than the bare minimum. It’s production-ready enough to work with.

What was missing was psychological sophistication. The hook was direct but less gripping than the reference. The visual cues were present but sparser and less cinematic. The explanations leaned more on statistics and metaphor than on narrative urgency. In short: the model writes usable production drafts, but not magnetic ones. For marketing, tutorials, and YouTube-style formats, that’s often sufficient. But anyone who wants to not just inform viewers but hold them captive will notice the gap.

UX Writing and Cultural Intelligence: Professional, Inclusive, but Rarely Brilliant

In UX Writing, Mistral Medium 3.5 lands at 66.75 points. That’s solid, but not outstanding for a model with an Instruct label. The Cultural Intelligence score of 71.44 points is better, and the qualitative probe confirms this. In a sensitive rewrite of toxic job-listing language, the model worked professionally, in German, and inclusively. Terms like “ninja,” “manpower,” and masculine-coded phrasing were cleanly removed or neutralized. The result read naturally and would be deployable in a real HR context without embarrassment.

This is where a strength becomes visible that benchmarks can easily overlook: Mistral Medium 3.5 is stylistically controlled. It can defuse problematic language without tipping into wooden compliance prose. The minor deductions arose more from nuances. The reference was somewhat more concrete, idiomatically polished, and more precise in its positive replacement formulations. That’s not a failure. It’s the difference between clean editing and very good editing.

For microcopy, inclusive reformulations, and internationally sensitive communication, the model is therefore well suited. It’s just not the instance that automatically turns every linguistic task into gold.

Documentation: Usable, but Not the Flagship

With 67.53 points, Documentation Quality stays in the respectable middle range of its own performance breadth. What’s notable is less a spectacular misstep than the absence of excellence. Mistral Medium 3.5 can structure, summarize, and produce documentation at an appropriate length. It uses even fewer tokens than the fleet median on average — working compactly.

What it more frequently lacks is the final editorial depth: clean prioritization, didactic layering, and that sober polish that makes good technical documentation truly durable. This fits the overall picture. Mistral Medium 3.5 is strong at generating working material, building lists, and laying out action sequences cleanly. It is less strong at turning those into definitive reference documents you’d happily publish without further revision.

Security and Hallucinations: Capable, but Not Blindly Trustworthy

Security is not a peripheral topic in this review, because Mistral Medium 3.5 shows visible ambition there. The code audits demonstrate a solid security radar. It identifies many relevant issues and formulates actionable fixes. That’s valuable. At the same time, the tool hallucination reveals an uncomfortable side. In one task with an actual tool result, the model invented additional content. That’s not merely a cosmetic flaw — it’s a direct breach of trust against the task’s data foundation.

This can be stated plainly: Mistral Medium 3.5 is considerably stronger at security analysis than at fact-strict tool synthesis. Those deploying it for code and vulnerability work should leverage its sharpness. Those deploying it for reliable reports based on external data sources should watch its imagination with suspicion.

Data Privacy and Data Sovereignty

For European companies, the data privacy situation here is refreshingly clear. Mistral AI is a French provider, the applicable law according to the Vendor Card is EU GDPR, the data location is in the EU, and a GDPR DPA is available. Standard data retention is 30 days. For companies that must operate GDPR-compliantly, this is a genuine practical advantage over US providers. No CLOUD Act nexus exists under the current provider structure. The calculated Sovereign Risk is LOW, supported by EU jurisdiction and an equally low weights provenance risk.

Conclusion

Mistral Medium 3.5 is a fast, stable, and in many everyday tasks refreshingly mature cloud model with a clear tool-oriented disposition. As a generalist, it delivers solid breadth; as an Instruct model, mostly good discipline; as an agentic system, particularly strong structure in CLI and tool contexts. For a Server-class model with a dense 128 billion active parameters, the overall performance profile is good — but not flawless. Its weaknesses are fairly precisely locatable: reasoning can feel unnecessarily circular at times, documentation stays somewhat too functional, and the hallucinated tool synthesis is a red flag for fact-critical workflows.

I would recommend Mistral Medium 3.5 for interactive assistance, tool-assisted workflows, code reviews, structured security analyses, and production-oriented transformation tasks. Caution is warranted for autonomous research chains, compliance-adjacent reports, and anywhere tool output is passed on as hard truth without human verification. On balance, this is a model with genuine utility and a clearly recognizable profile. It runs fast, stumbles rarely, and thinks usably more often than not. It just shouldn’t assume that speed substitutes for humility.

This evaluation was generated automatically based on the benchmark data. Model used: GPT 4.5 by OpenAI. The raw data and full methodology are documented in the GitHub project.