Qwen 3.5 9B (Unsloth)

With an overall score of 70.46%, Qwen 3.5 9B (UD-Q6_K_XL) delivers exactly the kind of performance you’d hope for from a dense 9B generalist in the Edge class: broadly capable, often surprisingly mature, but not without its scratches. As a Generalist without domain specialization, in the Size Class Edge and built on Dense architecture, this model doesn’t need to excel at everything. It needs to hold up across the full breadth. That’s mostly what it does, with notable strength in content adaptation and solid logic. The speed profile badge Interactive Content Adapter fits surprisingly well: this model is most at home when reworking, structuring, and shaping material into form — less so when asked to act as a security auditor or a strict format enforcer. Sovereign Risk: HIGH — the weights originate from Alibaba Cloud in China; while local deployment eliminates ongoing data transfer to a Chinese provider, the provenance remains relevant for sensitive procurement decisions.

Header Grades: Stability and Reliability

Architecture and Character: Generalist with the Handbrake On

The classification General, Thinking-Optional is not a footnote here — it’s the key to understanding this model. Qwen 3.5 9B (UD-Q6_K_XL) is not a coder’s scalpel and not a hardwired reasoning engine. It’s an all-rounder that in principle supports an extended thinking mode, but ran deliberately in standard mode during the benchmark. That’s methodologically clean. It measures the behavior users actually get without special configuration.

That’s precisely what makes the result interesting. The model feels smarter than its 9 billion parameters might suggest across many tasks. It reasons cleanly enough, writes with control, and doesn’t stand out for wild hallucinations. At the same time, you can sense that the unactivated thinking mode remains a kind of safety net sitting unused in the closet. On tasks with multiple simultaneous conditions, security depth, or particularly long chains of reasoning, the base model goes far — but not always all the way.

Speed and Local Deployment

Qwen 3.5 9B (UD-Q6_K_XL) was evaluated locally on an Apple Silicon M4 with 24GB Unified Memory (Shared RAM/VRAM). For an Edge model, that’s the relevant stage. At 34.32 tokens per second, it’s fast enough to feel interactive, but not so fast that its occasional latency spikes can be ignored. The Interactive Content Adapter badge says, at its core: this model is more the diligent editor than the frantic command-line sprinter. It suits rewriting, structuring, documentation work, and longer text tasks in dialogue. For genuinely time-critical agent chains with tight reaction windows, the pace feels only conditionally comfortable.

The hardware context matters. A 9B Dense model in high-quality quantization still sits in a reasonable zone on the test system. The swapping risk for large chunks remains limited, but it hasn’t disappeared entirely. The single timeout across 43 tests is therefore not a cosmetic flaw but a small warning light: the model runs appropriately for its class, but without much thermal or memory headroom.

Token economy is a positive. Across all budgeted modules, Qwen 3.5 9B (UD-Q6_K_XL) stays within expected bounds. No module runs away textually. For a local model, that simply means: no unnecessary self-extension, and therefore no artificially inflated wait times.

Reasoning and Logic: Competent, Not Majestic

At 70.73% in logical reasoning, Qwen 3.5 9B (UD-Q6_K_XL) lands in a respectable zone. The qualitative evaluation reveals a model that mostly understands the actual logic in core tasks. In the metacognition protocol on the classic guard puzzle, it arrives cleanly at the correct solution, explains the double negation correctly, and stays linguistically precise. That’s the good news.

The less good news is stylistic and structural. The model tends to introduce alternative phrasings or side paths and then fail to resolve them elegantly. In the protocol, it essentially says: this is where it gets complicated. Moments like that aren’t a reasoning crash, but they cost trust. If you’re explaining a logical problem, you shouldn’t be furrowing your brow mid-sentence.

The fact that the benchmark did not activate the optional thinking mode is central here. This model can in principle reach deeper. In standard mode, however, it shows workmanlike rather than luxurious reasoning. For everyday logic, structural synthesis, and reasoned responses, that’s often sufficient. For sensitive multi-step decisions where a wrong intermediate step poisons everything downstream, answers should be verified. Qwen thinks tidily. It does not think infallibly.

Code Quality and Security: Capable, but with Blind Spots

At 70.8% in Code Quality, Qwen 3.5 9B (UD-Q6_K_XL) presents as a serviceable technical generalist, not a forensic security specialist. The qualitative protocol is revealing on this point. In a PHP security analysis, the model reliably identifies the major issues: SQL Injection, plaintext passwords, Path Traversal, session problems, IDOR, CSRF, Header Injection. That’s not trivial. Many small models fail to cleanly name even the obvious holes.

The catch lies in depth and prioritization. The model finds 13 vulnerabilities, while the reference standard lists 19. Six missing findings in a security audit are not an academic gap — they’re a real risk factor. On top of that, Qwen rates several critical findings too mildly. When an IDOR case can potentially lead to admin takeover, “high” is simply too polite. In security, false restraint is not a virtue.

The fixes also tend to stay at the conceptual level. That’s formally permissible, since the task allowed short code snippets or concepts. In practice, though, it means the model usually says correctly what should be done, but less precisely how to implement it directly. For experienced developers, that’s acceptable. For teams hoping for near-copy-and-paste assistance, it’s too thin.

Also notable is a minor formatting weakness in the detail. In the security table output, a small Markdown error appeared in the column definition. Not a disaster, but symptomatic. This model is reliable enough for structured technical work, but not equipped with the unrelenting formalism of a specialized code model.

The security profile is ultimately split: good detection of the main issues, noticeable drop-off in completeness, severity calibration, and exploit depth. Usable for everyday code review. Not sufficient for serious security sign-offs without human cross-checking. Anyone deploying this model as an AppSec instance is confusing a flashlight with a floodlight.

Content Transformation: Its Strongest Card

At 80.73%, Qwen 3.5 9B (UD-Q6_K_XL) delivers its strongest benchmark discipline in Content Transformation. That’s no coincidence — it aligns cleanly with its speed badge. The moment raw material needs to be reworked, recast into a new form, or editorially charged, the model feels alert, organized, and surprisingly on-target.

The qualitative protocol on converting a dry 2FA tutorial into a YouTube-ready video script shows exactly this character. Qwen identifies the missing elements concisely and correctly, builds a complete time structure from 00:00 to 05:00, integrates screen annotations, B-roll, music cues, hooks, pattern interrupts, a CTA, and even a precisely described Easter egg. Above all, it does so not mechanically, but with a feel for format logic. The result is described by the Judge as “production-ready.” That verdict doesn’t come cheaply.

The technical coherence is also noteworthy. The 2FA steps remain accurate, the dramaturgy is realistic, and the language stays clearly German and suitable for spoken word. This is precisely where the strength of a good generalist shows: not maximally brilliant in a single discipline, but very strong at holding together requirements from content, tone, structure, and medium simultaneously.

The performance isn’t entirely without blemishes. The model tends toward over-delivery. In the example, the script grows beyond the desired word count because Qwen rolls out details generously rather than cutting hard. In content work, that’s usually the more forgivable sin — you’d rather cut than hunt for missing dramaturgy after the fact. Still, it remains a form of discipline problem. Anyone with tight production constraints needs to set boundaries clearly.

UX Writing and Documentation Quality: Useful, but Not Always Sharp

The module scores of 65.55% for UX Writing and 65.7% for Documentation Quality mark the zone where Qwen 3.5 9B (UD-Q6_K_XL) is reliably helpful without setting any highlights. That fits the overall picture. The model can phrase, explain, and structure. What it more often lacks is the final editorial sharpness.

The findings from the other protocols support this picture. Qwen generally writes cleanly, with grammatical stability and task orientation. It doesn’t drift aimlessly, mostly stays in the correct language, and produces no conspicuous fabrications. At the same time, its writing is not always as precisely condensed as one would expect for excellent documentation or pointed UX microcopy. It tends to explain solidly rather than distill. That’s a difference you notice immediately in practice when space suddenly becomes expensive.

For internal documentation, first drafts, rephrasing, and comprehensible intermediate steps, the model is well suited. For final user-facing copy with high demands on rhythm, tonal economy, and product polish, human editing remains the better option.

Cultural Intelligence: Decent, but Not Quite Elegant

In the Cultural Intelligence category, Qwen 3.5 9B (UD-Q6_K_XL) scores 65.6%. That’s not a failure, but it’s not a sign of particular cultural fine-tuning either. The protocol on detoxifying a toxic job posting illustrates the point well. The model removes problematic terms, writes consistently in German, converts aggressive phrasing into more professional language, and attempts to establish gender inclusion. The substance is largely correct.

But the elegance of the best solutions is missing. Rather than genuinely neutral nouns, Qwen sometimes falls back on masculine-coded terms and tries to smooth that over with an addition like “regardless of gender.” That’s functional, but not state of the art. You sense the good intent alongside a slight uncertainty in execution.

There’s also a tendency toward expansion. Where the task calls for a concise rewrite, the model adds its own accents and runs longer than necessary. In sensitive, inclusive communication, brevity is often not a side issue but part of the professionalism. Anyone who needs extra hedging loops to secure diversity language hasn’t quite internalized the core.

CLI and Tool Proximity: Sufficient, Not Outstanding

The CLI benchmark comes in at 77.78%, noticeably higher than some of the text-heavy disciplines. That suggests Qwen 3.5 9B (UD-Q6_K_XL) can generally handle structured technical instructions well. At the same time, the overall tooling profile remains limited, which is also reflected in the flanking tool values on the Leaderboard. This model is not a born orchestrator and not an obsessive command generator.

For simple shell tasks, pragmatic commands, and comprehensible technical steps, it’s capable. Where exact format matching, long agent chains, or hard fault tolerance are required, the comfort ends. That’s not a mark against its category. An Edge generalist is allowed to think smaller. You just shouldn’t push it into a role built for larger or more specialized models.

Privacy and Data Sovereignty

For this model, the most important point is the distinction between weight provenance and actual operation. The weights originate from Alibaba Cloud in China, and the calculated Sovereign Risk is therefore HIGH. The reason is not speculation but jurisdiction: Alibaba is subject to Chinese law, including PIPL, CSL, DSL, and the security-policy-relevant environment of the National Security Law. For European companies, cloud deployment via Alibaba would constitute a third-country transfer risk without an EU adequacy decision, even though a GDPR DPA is reportedly available per the vendor card and Alibaba offers regional data centers worldwide. The publicly stated retention period for API data is not clearly specified.

In the local deployment tested here, the operational privacy risk is substantially reduced, since no prompt data is sent to a cloud provider. For procurement and governance purposes, however, the provenance of the weights remains a valid decision factor. Anyone operating in strictly regulated environments should document exactly this distinction carefully: local execution is fine; the origin of the weights is nonetheless not neutral.

Conclusion

Qwen 3.5 9B (UD-Q6_K_XL) is a surprisingly mature Edge generalist with a clearly recognizable editorial talent. It shows its best side where content needs to be reworked, condensed, restructured, or translated into media-appropriate formats. Reasoning is good enough for many practical cases, code quality is solid, CLI work is serviceable. The weaknesses sit where completeness and uncompromising precision matter: security depth, severity calibration, absolute format precision, and long response time outliers. Across all tests, no notable hallucinations — the model prefers to invent too little depth rather than false certainty.

For productive deployment, that means: a very good model for local knowledge work, content transformation, document processing, general assistance, and technical first drafts. Less suited as the sole instance for security analyses, strictly supervised agent workflows, or tasks where every missing point is costly. The Apache-2.0-licensed Open Weights come from Alibaba; in local deployment, the cloud-specific privacy risk is significantly reduced, but the weights’ provenance remains a governance issue due to Chinese jurisdiction. Qwen 3.5 9B (UD-Q6_K_XL) is therefore no hero for every battle, but a very capable laptop all-rounder with a distinct profile. And in this class, profile is worth more than another smooth average.

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.