Qwen 3.5 4B (Unsloth)

With an overall score of 68.56 percent, Qwen 3.5 4B (UD-Q6_K_XL) delivers exactly what you’d expect from a dense Generalist in the Nano class with 4.0 billion parameters: more breadth than brilliance, more discipline than depth. The Speed Profile Badge Interactive Tool Expert fits surprisingly well. This model responds quickly enough for real dialogues while staying closer to the point across multiple disciplines than its size would suggest. Sovereign Risk: HIGH — the weights originate from Alibaba Cloud in China; local deployment eliminates ongoing data transfer to the provider, but the provenance remains a real sovereignty factor due to Chinese jurisdiction and the NSL.

Header Grades: Stability and Reliability

Performance and Runtime

Qwen 3.5 4B (UD-Q6_K_XL) is a local Open Weights model and was benchmarked on an Apple Silicon M4 with 24GB Unified Memory (Shared RAM/VRAM). Its official generation speed is 51.27 tokens per second. That’s not spectacular for this weight class, but very usable. The Interactive Tool Expert badge means, in plain terms: this model is geared toward conversational, direct workloads rather than heavy batch processing. It’s not meant to spend half the night grinding through a document in the background — it’s meant to respond promptly to concrete requests.

The architectural perspective matters here. Qwen 3.5 4B belongs to the Thinking-Optional category. CrucibleMark deliberately does not test this mode with Extended Thinking enabled. What’s being measured is the behavior a typical user gets without any special switches. If response times don’t always feel featherlight despite decent tokens-per-second, that’s not a measurement error — it’s part of the model’s character. It visibly doesn’t process like a bare command receiver; there’s somewhat more internal overhead. For a Nano model, that’s respectable. For time-critical tasks, it’s only limited consolation.

On the positive side: the model behaves token-economically. No module exceeds the expected verbosity range. In the CLI area it sits at 175 tokens versus a fleet median of 211; in Code Quality at 1876 versus 2112; in UX Writing at 1235 versus 1271. Only Documentation Quality and Content Transformation run slightly above, but nowhere near wasteful verbosity. For a local model, this is primarily a latency advantage. Qwen doesn’t write unnecessarily much. It only writes too much where it loses track of the constraints.

The Architectural Frame: Generalist, Nano, Dense

The General, Thinking-Optional classification is not a footnote for this model — it’s the key to a fair assessment. As a Generalist, it must compete across the full task spectrum: code, text transformation, UX, logic, documentation, cultural sensitivity. As a Nano model, no miracles in world knowledge or deep logic should be expected. As a Dense transformer, all 4.0 billion parameters are always active. There’s no MoE trick that looks large on paper while working smaller in practice. Whatever capacity is present here is genuinely the entire capacity.

Measured against exactly that, the result looks respectable. Qwen 3.5 4B (UD-Q6_K_XL) is not a mini genius. But it’s also not one of those small models that get nervous at the third constraint and then produce nothing but formatting dust. Its real strength is not excellence but a fairly robust all-round profile. For this class, that’s worth more than some benchmark fetishists care to admit.

Code Quality and Security: Usable, but Without Forensic Instinct

The hardest truth sits in the security section. In the Code Quality evaluation, the model reaches 63.5 percent, and that describes its character fairly precisely. It catches many obvious vulnerabilities but fails where security analysis shifts from catalog knowledge to systemic thinking.

The qualitative logs make this uncomfortably clear. In a PHP security analysis, Qwen identified 11 vulnerabilities while the reference standard found 19. That’s not a small gap — it’s a loss of threat picture. Missed items included reflected XSS, missing CSRF protection, debug exposure, missing expiration for reset tokens, and problematic database credentials. More critically, an IDOR-like account takeover chain was deprioritized as low severity, while the reference standard classified it as critical. That’s precisely where “can enumerate security terms” separates from “understands attack surfaces.”

The fixes are also often only half-complete. === instead of loose comparison is mentioned, but hash_equals() as a timing-safe comparison is absent. For path traversal, the whitelist appears, but realpath() validation against allowed directories does not. This is not a total failure. It’s the kind of answer that passes as a useful first-pass triage in a development team and then gets sharpened by the senior security engineer.

To its credit: form and language hold up. The model delivers a usable table, stays in German, and works in a structured manner. That doesn’t rescue the section, but it makes it productively usable. For simple audits, training material, or as a second pair of eyes on low-risk code, that’s fine. For real security sign-offs, making this model a gatekeeper would be negligent.

Logic and Reasoning: Surprisingly Solid, but Needs Room to Run

In Logical Reasoning, Qwen 3.5 4B (UD-Q6_K_XL) lands at 67.55 percent. That sounds unimpressive. For a 4B model, though, it’s more than a shrug. In the qualitative logs, it solves classic logic puzzles correctly, including clean inference and alternative reasoning paths. On the guard riddle, it formulates the right question, explains the double inversion coherently, and arrives at the correct conclusion. The Judge’s main criticism is the presentation: too long, somewhat cumbersome, occasionally not quite as crystalline as the reference.

That’s a recurring pattern. Qwen doesn’t think sloppily, but it doesn’t think elegantly. The answer has the slightly overloaded quality of a good student who genuinely understood what happened but hasn’t yet learned how to get to the point without background noise. That’s far more sympathetic than pseudo-confident nonsense. It just costs time.

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 is partially correct — the score deduction results from format non-compliance, not from reasoning errors. For comparison: in the tag-free reasoning_5* tests, the model achieves an average score of approximately 67.55 percent, consistent with its general reasoning level. CrucibleMark deliberately evaluates native zero-shot instruction compliance as a real-world characteristic. This deduction is methodologically intentional.

The single timeout episode in this module is also not random noise. Reasoning is the area where the Nano boundary becomes most visible. When tasks combine many conditions, format requirements, and longer execution, the model starts to strain. It remains useful. It just doesn’t remain composed.

Content Transformation: Strong at Restructuring, Weak on Word Limits

At 73.21 percent, Content Transformation is one of the more encouraging areas. The model can reshape texts, defuse toxic language, professionalize tone, and transfer working material into new formats. This shows most clearly in the inclusive rewrite of an aggressive job posting: gender-neutral language, toxic combat terms removed, professional tone established. The Judge’s main criticism there is that Qwen writes correctly rather than warmly. The reference feels more inviting, more idiomatic, more human. Qwen reads like HR after a night with compliance documents. That’s safer than embarrassing, but not exactly a charm offensive.

On more complex transformation tasks — such as a video script with timing, hook, visual cues, and an Easter egg — the model shows genuine production proximity. It names missing elements, sets timestamps, provides on-screen cues, and understands YouTube mechanics better than you’d spontaneously credit a 4B model with. The Judge explicitly calls the result production-ready, with minor weaknesses in cue precision and structure.

And that’s precisely where the discipline gap strikes. In two tasks in this module, the model significantly exceeded explicit word limits. That’s not a cosmetic error — it triggers an automatic point deduction by the system.

In one task in the Content Transformation area, the model exceeded the explicit word limit of 250 words by 27 percent. The system applied an automatic deduction of 20 percent, or 11.80 points, on the achievable task score. The substantive quality of the answer is irrelevant at that point. The penalty applies regardless.

In a further task in the Content Transformation area, the model exceeded the explicit word limit of 900 words by 25 percent. The system again applied an automatic deduction of 20 percent, or 17.60 points. Here too: whether the content was good has no bearing on the penalty.

The length problem is not an isolated outlier. Across multiple tasks in the Content Transformation area, the model shows a consistent pattern: when simultaneous constraints of language, length, and format are present, it drops the word limit as the first condition. That’s a classic Nano weakness. Anyone deploying this model for editorial or marketing workflows should either build in hard post-checks or frame the word count constraint more aggressively in the prompt. Otherwise Qwen delivers a usable text that grows past the brief like ivy on a wall.

Documentation Quality: Solid, Factual, Somewhat Expansive

In Documentation Quality, Qwen reaches 68.64 percent. Not a triumph, but a stable showing. The model writes coherently, in a structured manner, and with a visible commitment to completeness. The slight token overhead of 1.13 versus the fleet median suggests it prefers to lay out explanations a bit more broadly without tipping into rambling.

The overall impression from the logs: Qwen documents like a conscientious teammate who doesn’t want to leave anyone stranded. That’s often useful. It becomes laborious where conciseness is the goal. For internal documentation, onboarding material, or comprehensible first drafts, it’s quite suitable. For highly compressed technical references, it lacks the final editorial sharpness.

UX Writing and Microcopy: Surprisingly Sure-Footed

At 71.75 percent in UX Writing, Qwen 3.5 4B (UD-Q6_K_XL) is one of the more pleasant surprises in this benchmark. Small models often confuse microcopy with mini-marketing or deliver sterile button labels that look like they fell out of a form graveyard. Qwen does better. It stays close to the task, writes concisely enough, and doesn’t systematically miss the tone.

This isn’t a poetic gift. It’s something more useful: functional language discipline. For interface text, error messages, small reformulations, and all the inconspicuous copy by which products are judged daily, the model is more credible than its size would suggest.

CLI Benchmark: Pragmatic Rather Than Brilliant

The CLI area comes in at 72.78 percent. That fits the overall picture. Qwen is no terminal virtuoso here, but it’s genuinely workable. The low token consumption shows it doesn’t inflate commands unnecessarily. For simple to moderate shell tasks, that’s exactly right. The model tends to say what needs to be done rather than talking at length about it.

Caution is still warranted for agentic workflows. A Nano model with sporadic timeouts and limited depth of focus is not a tool to which you should entrust unsupervised long multi-step sequences. As a local assistant for individual commands, parsing help, or everyday Linux tasks, however, it’s well suited.

Cultural Intelligence: Polite, Competent, but Not Particularly Warm

The score of 64.6 percent in Cultural Intelligence looks a bit lean at first glance, but the logs read better than the number. The model reliably defuses toxic or exclusionary language, stays stable in German output, and mostly hits inclusive phrasing adequately. In the concrete example of a problematic job posting, it consistently removes terms like “ninja,” “kill the competition,” and other macho reflexes. That’s not trivial. Many models only smooth the surface here while leaving the underlying attitude intact.

The catch: Qwen writes culturally correctly, but not with particular sensitivity. The Judge flags a somewhat traditional, requirements-focused tone and a lack of warmth in the invitational quality. Put differently: the model knows what you shouldn’t say anymore. It doesn’t always know yet how good, modern outreach sounds elegant in German. That’s a distinction HR departments and communications professionals will notice immediately.

Data Privacy and Data Sovereignty

For this locally operated Open Weights model, the decisive point is not a cloud privacy section but the origin of the weights. The calculated Sovereign Risk is HIGH. The reason is provenance: the developer is Alibaba Cloud in China — a company under Chinese jurisdiction with relevance to PIPL, CSL, DSL, and the politically significant National Security Law operating in the background. For users in Germany and Europe, this is considerably less sensitive in local deployment than with direct API usage, since no ongoing prompt data flows to the provider. The origin of the weights nonetheless remains a sovereignty factor that regulated environments should not simply wave away.

Conclusion

Qwen 3.5 4B (UD-Q6_K_XL) is a small model with a surprisingly mature bearing. It reaches 68.56 percent, works in a mostly structured manner, stays token-economical, and delivers a convincing everyday breadth for a Nano-class Generalist. Its best sides show in UX Writing, CLI proximity, solid text transformation, and usable standard reasoning. Its weaker sides lie where depth, prioritization, and hard constraint adherence are required: security analysis remains incomplete, word limits break repeatedly, and tail latency is too high for longer reasoning tasks. Across all tests, no notable hallucinations — the model prefers to invent too little context rather than ruin itself with fabrication.

The weights come from Alibaba Cloud under the Apache 2.0 license. For local use, that’s practical and clearly more favorable from a sovereignty standpoint than a cloud call, even if the elevated provenance risk from Chinese origin persists in the background.

The recommendation is correspondingly clear: this model is suited for local everyday tasks, writing assistance, light development support, microcopy, reformulations, and simple agentic subtasks with oversight. Anyone needing precise security assessments, hard format compliance without rework, or deep, fast multi-step reasoning should buy up a class. Qwen 3.5 4B (UD-Q6_K_XL) is not a scalpel. But for a pocket knife, it cuts remarkably cleanly.

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.