Qwen 3.5 35B-A3B (Unsloth)

With an overall score of 74.72%, Qwen 3.5 35B-A3B Q4_K_XL delivers a surprisingly cohesive package for a generalist in the Workstation class. The model is clearly trimmed for direct execution as an instruct system, supports Extended Thinking, but ran here in standard mode. That is precisely what makes this model’s character interesting: fast, disciplined, often accurate, but not free from the minor sins of a system that prefers pragmatic answers over pedagogical polish. Sovereign Risk: HIGH — the model vendor is based in China; in cloud deployments, Chinese legal jurisdictions would become relevant, even though this particular test subject runs locally.

Header Grades: Stability and Reliability

Architecture and Classification

Qwen 3.5 35B-A3B Q4_K_XL is a generalist with an instruct focus, optional thinking, and a multimodal design. This combination sets the evaluation frame fairly clearly. You should not expect a pure coding specialist, but very much a broad all-round model that responds cleanly to direct instructions. The Multimodal tag is simultaneously a strength and a limitation of what can be said here: the benchmark only tests the text side. Image understanding is part of the model’s identity but was not exercised in this run.

More important is the second axis: Workstation class and MoE architecture. The spec sheet lists 35 billion parameters, but only 3 billion are active per token. This is not a detail for spec-sheet enthusiasts — it is the core of the model. Qwen does not play with the raw mass of a dense 35B transformer but with a selective expert logic. Performance therefore needs to be calibrated against the active 3B. And that is precisely where this model feels remarkably mature. It does not respond like an inflated colossus but like a very well-trained specialist in a tailored suit.

Speed and Runtime Profile

The Speed Profile Badge reads Real-Time DevOps Expert. That is an apt shorthand for this model’s character: not merely fast, but designed for direct, productive interaction. In practice, this means the model does not feel like a batch writer that takes a deep breath before mumbling away, but like a system that steps up immediately for shell, code, and structured work orders.

On the local reference system with Apple Silicon M4, 24 GB Unified Memory (Shared RAM/VRAM), that is more than an academic footnote. A Workstation model with 35B total parameters quickly brushes up against real memory limits. That Qwen remains this responsive in this quantization is a genuine achievement of the MoE design and the Q4 trade-off line. The test system did not fall into the typical VRAM pitfalls of many larger Open Weights candidates. Those working locally get a model that looks like serious weight but behaves more like a disciplined mid-class athlete than an unpredictable heavy hauler.

Then there is token economy. Across all modules, the model stays below the fleet median, sometimes significantly. It behaves token-economically — no module exceeds the expected verbosity envelope. For a local model, this is not a cosmetic virtue but directly perceptible practical quality: less textual filler, less wait time, less risk of long responses sprawling unnecessarily.

Code Quality: Technically Solid, but Not Flawless

In the code and security domain, Qwen 3.5 35B-A3B Q4_K_XL shows one of the more encouraging sides of its architecture. It can identify vulnerabilities, deliver clean Markdown tables, and name concrete fixes without losing itself in self-congratulatory security prose. That sounds obvious but is by no means a given in local models of this class.

The security audit excerpt makes the direction clear. The model identifies central vulnerabilities such as SQL Injection, Session Fixation, Path Traversal, weak token generation, type juggling, CSRF, and IDOR. It also provides technical countermeasures that are largely sound. Particularly positive is that it addresses the required implicit vulnerabilities rather than merely collecting the obvious labels from the source code. This capability separates a usable audit model from a mere pattern matcher.

The weakness lies not in technical instinct but in coverage. In the specific audit, the model stopped at 11 rather than 19 relevant findings. That is not a minor point. Missing eight gaps does not constitute a complete audit — it constitutes a solid preliminary analysis. The severity ratings are also a concern: Path Traversal was rated too low, as was an insecure cookie authentication. In security work specifically, incorrect severity ratings are dangerous because they cause teams to work in the wrong order. An overlooked bug is annoying. An underestimated bug is, in the worst case, far more costly.

The same pattern appears in the fixes. Much is correct; some things stop half a step short of best practice. The model recommends strict operators and input validation for loose comparisons but misses the timing-safe comparison via hash_equals(). That is not a beginner’s mistake, but it is also not the confidence of a truly excellent security model.

On balance, Qwen is clearly usable here. For initial analyses, review assistance, and structured triage work, it is strong. For a final security sign-off without a human review, the precision is not yet sufficient. Put differently: a reliable sniffer dog, not a sole auditor.

CLI and Tool Proximity: One of Its Most Natural Disciplines

The numbers in the CLI domain and the badge tell the same story. Qwen 3.5 35B-A3B Q4_K_XL is built close to DevOps tasks — or at least exceptionally well-tuned for them. The model operates concisely, quickly, and with sufficient precision for operational text tasks around shell, structuring, and technical directives. The instruct character plays directly into its hands here. Where other models begin by explaining the philosophy of the command line, Qwen tends to deliver the tool.

This is not coincidence but an expression of the model category. An instruct system should comply, not preach. In this benchmark, that works remarkably well.

There is, however, a flaw that should not be glossed over. In a tool-use task, the model hallucinated content that did not originate from the retrieved tool result. The score was consequently capped by a hallucination cap. For content-critical tasks such as research, fact-bound reports, or agentic tool pipelines, this is a hard warning. Once external results are the source of truth, a model must not improvise. That is precisely where Qwen improvised in this case. This is not a creative slip — it is a small but real breach of trust.

Reasoning and Logic: Thought Through Correctly, Not Always Delivered Cleanly

For a Thinking-Optional model, the reasoning domain is particularly revealing. Extended Thinking was not activated here. What was evaluated was out-of-the-box behavior. And that is solid. Qwen argues in a structured manner, achieves usable to good results on logic tasks, and demonstrates that genuine analytical depth lies beneath the fast surface.

The specific metacognition excerpt on the guardian puzzle confirms this. The solution was correct, the derivation sound, the structure comprehensible. The model was able to resolve the case cleanly and explain the core logic. What was missing was less the correct answer than the final layer of didactic elegance. It does not name the abstract principle level, does not offer a particularly strong generalization, and remains functional rather than textbook-ready in its presentation. For users, that is often entirely sufficient. For a benchmark that also measures proximity to reference standards, it costs points.

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 reasoning errors. For comparison: in the tag-free reasoning_5* tests, the model achieves an average score of approximately 70.39%, 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.

There is also an unfortunate language error. In one reasoning task, the model ignored the explicit language instruction and responded in English. This is not a reasoning error but a weakness in instruction-following. In production environments with a fixed target language, this is not a minor infraction — it is a direct quality failure.

In the same task, this violation was additionally penalized by a rule-based mechanism. The instruction explicitly required German, yet the system flagged a Language Mismatch. The situation here is somewhat confusing, because the Judge evaluates the German final answer as compliant, while the automated language check penalizes the English content within the <thought> tags. What matters most for the reader is this: the model can reason cleanly but visibly loses discipline under nested format and language requirements. Thinking-Optional models in particular sit between two worlds here. They have the depth, but not always the outward rigor.

Content Transformation: Strong at Restructuring, Vulnerable Under Strict Guardrails

In the Content Transformation domain, Qwen 3.5 35B-A3B Q4_K_XL shows much of its actual talent. The model can not only rephrase raw material but functionally rearrange it. The video script test is a good example. Qwen delivers a complete, German-language, well-structured version with hook, troubleshooting, CTA, B-roll, music cues, and screen annotations. Many models fail here on completeness or rhythm. Qwen does not. It builds the scaffold cleanly and keeps the production purpose in view.

The weaknesses are craft-level issues but are genuinely relevant for real media work. In the script in question, time segments overlap. That may sound like a marginal detail, but in a production document it is simply an error. Anyone working with timings cannot issue contradictory cues. Similarly, the pattern interrupt was placed far too late in the sequence. The model understood the concept but placed it strategically incorrectly. That is typical of a good, not excellent, transformation model: it knows the vocabulary of dramaturgy but does not always hit the beat.

There is also a hard constraint violation. In a task within the Content Transformation domain, the model exceeded the explicit word limit of 250 words by 31%. The system applied an automatic deduction of 20%, or 12.40 points, to the achieved partial score. The substantive quality of the response is therefore irrelevant — the penalty applies regardless. Anyone working with editorial guidelines, social copy limits, or UI text budgets should take note here: Qwen writes well, but not always concisely enough when multiple guardrails apply simultaneously.

UX Writing and Microcopy: Precise Enough, but Without the Final Polish

In the UX domain, the model’s instruct nature again works in its favor. It formulates with focus, does not drift into marketing fog, and generally keeps functional tasks on track. For microcopy, that matters more than literary flair. Buttons, notices, system messages, and short user dialogs benefit more from plainness than from flashes of inspiration.

What distinguishes Qwen here is a workable balance of brevity and clarity. Since the model also remains token-economical, there is rarely the impression that it wants to inflate microtexts into mini-essays. That is a genuine advantage in practice. What you get less of is the particularly fine emotional calibration that top-tier models sometimes command. Qwen often sounds competent but not always maximally nuanced. It is the voice of a good product manager, not an outstanding brand writer.

Documentation Quality: Workable, but Not Luxurious

In the documentation domain, Qwen lands in solid middle ground above what many local models deliver. It can structure, explain, and expand longer responses in a controlled manner without losing the thread. The large context window is a relevant asset here, even if this benchmark does not push extreme long-context scenarios. The model generally gives the impression that it can manage large volumes of text adequately, as long as the test system is not put under pressure by unusually long context chains.

The weakness lies less in structure than in depth. Qwen documents adequately but not luxuriously. It does not build particularly elegant didactic dramaturgy, does not always abstract far enough, and delivers more of a solid working draft than a document you would want to drop straight into a manual. That is not a criticism — it is a positioning statement. For internal documentation, technical summaries, and first drafts, it is more than sufficient. For publishable final versions, revision remains advisable.

Cultural Intelligence: Competently Localized, but Not Quite Fine-Tuned Stylistically

Qwen’s strength in the Cultural Intelligence domain is its robust baseline competence. It stays in German, removes toxic or inappropriate phrasing, and understands the functional core of the task. When rewriting a problematic job posting, the model successfully replaced aggressive, gender-coded language with more neutral and professional formulations. That is more than mere language polishing. It shows that the model can read social registers at least at a practically useful level.

What was missing was the finer country-specific tonal register. The Judge rightly noted that the formulations, while correct and professional, came across as somewhat bureaucratic and less inviting than the reference. That is typical of a model that handles compliance better than charm. It offends no one. It does not automatically inspire anyone either. For HR-adjacent or brand-sensitive communication, that is a distinction that matters.

Hallucinations and Safety Profile

Hallucinations deserve a dedicated look with this model, particularly because the rest of its performance appears so controlled. Across broad text and structure tasks, Qwen tends toward caution rather than fabrication. Yet the documented tool-use case shows that this caution is not absolute. Once external tool results are supposed to be the basis of a response, the model can add content it cannot substantiate. In an agent pipeline, that is genuinely dangerous. A system that formulates shell instructions well and then creatively “improves” tool outputs is like a meticulous accountant with a tendency to invent receipts.

For security, a similar dual picture applies. The model identifies many relevant vulnerabilities and proposes largely sensible measures. At the same time, its coverage is not complete enough to deploy it as a sole authority. Qwen is therefore a good assistance model for technical security work, but not one to which you should hand the final red stamp.

Data Privacy and Data Sovereignty

For this specific test subject, there is no acute cloud data exposure because the weights run locally. What matters therefore is weights provenance, and that is comparatively uncritical here: Apache-2.0-licensed, publicly available weights in an Unsloth quantization, with no mandatory contact with any external provider infrastructure. The stated provenance risk is accordingly low.

Conclusion

Qwen 3.5 35B-A3B Q4_K_XL is one of those models you come to respect faster than the spec sheet might lead you to expect. As a local Open Weights system on the test hardware, it combines high practical speed with a surprisingly broad competence base for only 3 billion active parameters. Code, CLI, structured transformation, and general instruction-following are often convincing. Token economy is good, stability is flawless, and the MoE architecture shows its real appeal here: it saves weight without immediately sacrificing substance.

The limits are nonetheless visible. Security analyses are usable but not complete. Reasoning is often substantively sound but loses format discipline under meta-requirements. Content transformation is largely strong but stumbles on hard word limits and individual production details. And the documented tool hallucination case is not a peripheral error — it is a genuine warning for agentic or fact-critical deployments.

My recommendation is therefore clear: excellent as a local all-round assistant for developers, technical writers, and productive knowledge work with human final review. Less suitable as an autonomous authority for security sign-offs, tool-bound factual outputs, or strictly regulated language and format processes. Weights provenance is pleasingly clean for local use. Qwen does not hallucinate never — but often enough it delivers exactly what a good local model should: plenty of substance, little drama.

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.