Gemma 4 31B Ortenzya Creative Wordsmith (Thinking)

With an overall score of 72.56%, Gemma 4 Ortenzya Creative Wordsmith 31B (NVFP4) presents a remarkably ambivalent profile: strong in logic, inconsistent in instruction-following, and alarmingly weak in operational stability. The Speed Profile Badge reads Unusable DevOps Expert, and the name is not polemic — it’s a diagnosis: this model thinks thoroughly, works intelligently in places, but behaves throughout the benchmark like a talented author who consistently misses deadlines. As a generalist Workstation model with 30.7 billion active parameters in a dense architecture, broad competence is a reasonable expectation. That is precisely why the gap between reasoning ability and operational reliability weighs so heavily here.

Header Notes: Stability and Reliability

Architecture and Expectations

The pre-assigned category fits surprisingly well at its core, though not without friction. Reasoning and Thinking are not label-washing: this specific run was conducted explicitly in Thinking mode, and it shows. Responses are often longer, more step-by-step, and argumentatively structured rather than simply emitted. For logic tasks in particular, this is a strength, not dead weight.

At the same time, the model also carries the Instruct gene. It visibly wants to work through tasks, deliver tables, reconstruct structures, and formally satisfy requirements. But this does not hold consistently. In several instances, when faced with combined constraints on language, length, and format, it loses the precision one expects from a well-behaved instruct model. That is not a cosmetic flaw — it is a functional weakness.

Open-Weight and Multimodal set the utility context. This is a locally usable, freely available model based on the Gemma-4-31B line, with image capability in the base. Importantly, however, this benchmark is text-heavy. The multimodal capabilities are only visible here indirectly. What is being evaluated is therefore only part of what the model can do.

The Uncensored tag is both interesting and delicate. We are not talking about a harmlessly freer chat model, but a community variant with deliberate abliteration followed by style fine-tuning for creative writing. This explains the more natural tone and reduced tendency to refuse. It also explains why security, code, and strict enterprise behavior deserve particularly close scrutiny. Cutting guardrails out of a model does not automatically yield more capability. What you get, above all, is less braking.

Speed and Working Character

As a local model, Gemma 4 Ortenzya Creative Wordsmith 31B (NVFP4) ran natively on an ASUS GX10 / NVIDIA DGX Spark (GB10 Grace Blackwell Superchip, ~115 GB Unified Memory — no practical memory limit for tested model sizes). The Speed Profile Badge Unusable DevOps Expert describes the character accurately: not because the model lacks substance, but because its response speed — and especially its outliers — disqualify it from interactive technical workflows.

In operation, it does not move lightly — it moves heavily. The model is token-economical. No module exceeds the expected verbosity range. This matters, because the slowness here does not stem from uncontrolled verbosity. In Reasoning and Metacognition, it stays well below the fleet median of 1,207 tokens with an average of 782 output tokens; across all other modules it is consistently below as well. In other words: it does not talk too much — it just takes too long regardless. For local users, this is an uncomfortable finding, because the problem cannot be moderated away with shorter prompts.

The badge implies a DevOps use case. That is precisely where this behavior would be fatal. In an agent framework, review loops, or iterative CLI tasks, what matters is not only whether the answer is smart, but whether it arrives reliably. This model responds too often like an expert who needs “just a minute” — and then never shows up.

Reasoning and Logic

This is where the model shows its best side. A Reasoning score of 77.05% is strong for a Workstation model in this class, and the qualitative logs confirm the impression. On the classic two-guards puzzle, it delivers the correct question, cleanly explains both cases, and ends with a clear course of action. Not academically polished, but comprehensible. That is solid model work.

The style of its reasoning is notable. Gemma 4 Ortenzya Creative Wordsmith 31B (NVFP4) does not build brittle formal proof structures; it favors narrative derivations. The Judge rightly notes the absence of terms like “double inversion” or alternative formal formulations. In return, the explanation is directly readable. That is a fair trade, as long as the logic holds. Here, it does.

The catch lies elsewhere: a thinking model that visibly enjoys thinking must also deliver that thinking reliably. That is precisely what fails too often across the module. The substantive logic is convincing; the operational stability is not. One can respect the model’s reasoning and still deny it a production role. Both positions are factually correct here.

Code Quality and Security

At 67.44% in the Code Quality Audit, the model initially appears adequate. Reading the logs reveals a clearer picture: it identifies real vulnerabilities but misses too many of the subtle and structurally decisive ones. In a PHP security audit, it correctly identifies 13 of 19 vulnerabilities, including SQL Injection, plaintext passwords, XSS, IDOR, Path Traversal, Session Fixation, and Information Disclosure. That is not a total failure. In particular, it largely finds the five implicit weaknesses the prompt specifically targeted.

Nevertheless, an uncomfortable aftertaste remains. In security work, completeness is not a luxury — it is the core of the task. The model misses, among other things, a weak reset token, hardcoded secrets, missing CSRF protection, absent token expiration, and problematic database credentials. These are not bonus points. These are gaps a real attacker would walk straight through.

There is also a characteristic flaw of this fine-tune lineage: the response is formally clean but analytically too shallow. The Markdown table is there, the columns are correct, the language is proper. What is missing are attack chains, prioritization, and a clear statement that the code is not production-ready in its current form. The model describes security vulnerabilities but does not drive home their interplay forcefully enough. It is more a tidy auditor than an uncompromising incident responder.

In fairness, it should be noted that an uncensored creative-writing fine-tune was not built primarily for coding and security. But at this size class, that does not excuse everything. A 30.7B dense generalist at Workstation level should demonstrate more depth in this domain.

CLI and Tool Proximity

The CLI benchmark comes in at an excellent 93.67%. This is arguably the model’s real surprise. The Creative Wordsmith rework could easily have caused shell-adjacent precision and tool discipline to fray. It did not. On the contrary: the model delivers precise and actionable responses in this area.

The result should not be romanticized, however. A high CLI score means primarily this: when an answer arrives, it is often good. It does not cancel out the stability problems. For real tool-use pipelines, the verdict is therefore paradoxical: technically interesting, operationally risky. The distinction matters. A model can look like an excellent command-line assistant on paper and still be as frustrating in daily use as a server that silently drops every fifth SSH session.

Content Transformation: Style Present, Discipline Absent

This module is where the fine-tune character is most clearly felt. The score of 63.38% is not catastrophic, but it is too low for a model trimmed for natural writing. The logs explain why.

In a video script task, the model delivers many of the required elements: timestamps, production cues, a hook, a pattern interrupt, a CTA, even an Easter egg. The content is usable. But it responds in English when German was explicitly required. That is not a minor slip — it is a clear violation of the task condition. The model ignored the explicit language instruction and responded in English. In production environments with a fixed target language, this is a direct operational risk.

More damaging still: in one Content Transformation task, the model exceeded the explicit word limit of 900 words by 21%. The system applied an automatic deduction of 17.60 points, or 20%. The substantive quality of the response is therefore irrelevant — the penalty applies regardless. That is exactly how a benchmark must treat such violations. Treating a word limit as optional is not a failure of taste; it is a failure of craft.

The language failure is not an isolated outlier either. Across multiple tasks in the Content and Documentation modules, the model shows a consistent pattern: when faced with simultaneous constraints on language, length, and format, it drops the language requirement first. This is an uncomfortable finding, because it directly undermines the model’s instruct component. The creative impulse is there. The discipline disappears the moment multiple reins are pulled at once.

Documentation Quality

At 75.06%, Documentation Quality appears robust. Here too, however, a structural fault line lurks in the detail. In one documentation task, the model ignored the explicit language instruction and responded in English. This is not a technical slip — it is an instruction-following weakness. In product documentation, internal knowledge bases, or compliance texts, this is not merely careless; it is unusable.

Setting that aside, the model appears to work structurally well in documentation tasks. The score suggests solid organization, acceptable completeness, and clear presentation. But the English outlier pulls the entire module down, because it does not merely reveal a stylistic preference — it reveals unreliability on an explicit core requirement.

UX Writing and Cultural Intelligence

In UX Writing, the model reaches 71.05%. This fits its character surprisingly well. It can write readably, clearly, and accessibly, but does not always feel as finely calibrated as specialized microcopy models tend to be. Responses often have substance, but not always the last degree of tonal precision.

The Cultural Intelligence score of 71.72% is even more revealing. In a task rewriting toxic job-ad language, the model cleanly identifies the problematic terms, removes gender bias, and replaces aggressive formulations with professional alternatives. That is technically competent. The Judge’s main criticism is a lack of warmth, insufficient use of initiative as a positively framed value, and overall language that feels somewhat sterile and corporate. This hits the mark well: the model reliably sanitizes language, but not always elegantly. It disinfects where one might sometimes prefer to cultivate.

For a Creative Wordsmith fine-tune in particular, this is a small sting. The natural language is present; the emotional tonal modulation is not as pronounced as the name would suggest. The text sounds correct. It just does not always sound alive.

Data Protection and Data Sovereignty

No dedicated privacy section is needed here, as this is a local Open Weights model with no cloud provider involved in active use. What remains relevant is the provenance of the weights: the Weights Provenance Risk is rated MEDIUM. The base model originates from Google DeepMind under the Apache 2.0 license, but the specific variant is a community chain of abliteration, creative fine-tuning, and NVFP4 quantization by a single contributor without documented jurisdiction. For organizations, this is not a classic data protection issue, but it is a governance concern: free local use is permissible, but with heightened diligence regarding origin, security profile, and internal approval.

Conclusion

Gemma 4 Ortenzya Creative Wordsmith 31B (NVFP4) is a model with capability and loss of control. It reaches 72.56% and demonstrates genuine strengths in logic, CLI proximity, and structured writing. At the same time, it is devastatingly weak in operational stability, surprisingly disobedient with language instructions, and too incomplete in security analysis to be trusted without supervision. Across all tests, no notable hallucinations — the model prefers to invent little rather than ruin itself with wild fabrications.

The comparison with the second run of the same model is instructive: in standard mode it sits at 72.54%, virtually identical. Thinking mode lifts Reasoning slightly and also improves the CLI area, but pays for this with an even heavier runtime profile and no meaningful gain in the creative modules. That is the real point. More thinking does not make this model fundamentally better — it makes it, above all, more demanding.

I would recommend it for local, supervised work on logic tasks, rough drafts, freer writing, and individual technical support jobs where time is not a dominant factor. I would not recommend it for unattended agents, time-critical workflows, multilingual production environments with strict language requirements, or security tasks carrying real responsibility. This model is not a fraud. But it is not a rock either. More of a gifted free spirit with a tendency toward tardiness. In literature, that can have charm. In infrastructure, it has consequences.

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.