Qwen 3.8 27B Uncensored (Thinking)

With an overall score of 79.75 percent, Qwen 3.8 27B Uncensored is not an anarchic bastard with broken brakes, but a surprisingly controlled generalist in the Workstation class. The classification fits: a dense 27.8B model for broad tasks, locally deployable, tested in this run with Thinking mode enabled and thus explicitly calibrated for longer, more deliberate responses. Its Speed Profile Badge reads Unusable Tool Expert. That says less about intelligence than about temperament: strong in tool- and DevOps-adjacent tasks, but with an execution profile that too often becomes sluggish and erratic in direct workflows.

Header Scores: Stability and Reliability

These header scores are not minor footnotes. They cut across the solid overall score and define the model’s real character far more than a clean average does. Qwen 3.8 27B Uncensored can deliver on substance. It just fails to deliver reliably often enough. Anyone looking for a local model for agents, batch automation, or unattended tool chains will not find a Swiss watch here — more like a highly talented employee with excellent ideas and unacceptable punctuality.

Architecture and Expectations

The metadata captures the model fairly precisely. Qwen 3.8 27B Uncensored is classified as a Generalist, not a specialist. And that is exactly how it behaves. Its strengths do not lie in a single domain but in breadth: code auditing, logic, cultural rewriting, documentation, and tool context are all competent enough across the board to be taken seriously.

The Workstation class matters here. 27.8 billion parameters in a Dense architecture means all weights are active for every response. Unlike MoE models, where only subnetworks are active, the nominal size here directly corresponds to active capacity. Expectations can therefore be elevated. This model does not need to hide behind “pretty decent for its size.” It is supposed to deliver — and on substance, it frequently does, at mid-tier cloud level.

The sensitive point is the Uncensored label. According to the Model Card, this is an abliterated community variant based on Qwen 3.8 — not merely a looser fine-tune, but targeted interventions in the refusal direction of the weights. Such interventions visibly ruin some models. Here, pleasingly, that is not the case. The claimed capability loss of around two MMLU points seems plausible in the benchmark. Qwen 3.8 27B Uncensored does not behave like a gutted safety experiment but like an intact base model with the handbrake released. That deserves respect. It does not automatically make deployment wiser.

Then there is the Thinking mode of this specific test run. That is not a side note. Longer responses, more structure, and more elaborate derivations are intentional here and must be read positively, as long as they translate into quality. The model does this frequently. It just purchases that depth too often at the cost of poor operational stability.

Speed and Efficiency

The Speed Profile Badge Unusable Tool Expert is bluntly worded, but accurate. On substance, the model is suited for tool-adjacent work. Operationally, it is too uneven for it. In practice, that means: usable for individual, supervised specialist tasks; only with caution for fluid interaction or reliable agent chains.

On the local reference system ASUS GX10 / NVIDIA DGX Spark (GB10 Grace Blackwell Superchip, ~115 GB Unified Memory — no practical memory limit for tested model sizes), one thing stands out above all: memory is not the bottleneck here — the model’s own operational profile is. The test system sees a model that works more in batch mode than interactively, with outliers too massive to explain away.

On the positive side, there is the token economy. Qwen 3.8 27B Uncensored behaves with discipline across most modules. No module exceeds the expected verbosity envelope. For a Thinking model, that is noteworthy. It does not produce pointless walls of text but invests its length mostly where it is substantively warranted. That does not make the stability problems smaller, but it at least prevents the second classic failure mode of reasoning-heavy models: expensive and verbose at the same time.

Reasoning and Logic

In reasoning, Qwen 3.8 27B Uncensored shows its most mature side. The logic score of 78.43 percent is not a coincidence. In the available logs, the model works cleanly, in a structured manner, and with recognizable methodical discipline. On the classic two-guards puzzle, it formulates the correct counter-question, checks both cases consistently, and reliably arrives at the correct door. The Judge’s criticism targets pedagogical depth rather than logical correctness. That is a good way to lose points.

The tone of these responses matters. The model does not explain frantically but in an orderly fashion. Tables, numbered steps, intermediate verification: not a fireworks display, but it builds trust. In Thinking mode especially, that is a quality marker. Some reasoning models burn their additional depth in convoluted self-circling. Qwen 3.8 27B Uncensored mostly puts it to useful ends: clarity.

It is not entirely flawless. The Judge notes a slightly imprecise explanation on an alternative formulation of the puzzle. The result remains correct, but the reasoning loses some precision. That is precisely where this model’s limits become visible: it thinks soundly, but not always elegantly. Anyone seeking philosophical meta-justifications, particularly deep generalization, or didactic polish will get solid engineering work — not a great essay.

Code Quality and Security

In Code Quality, Qwen 3.8 27B Uncensored delivers one of its strongest performances. The score of 84.52 percent is earned. In the security audits, the model reliably identifies the decisive classes of modern web vulnerabilities: SQL injection in multiple variants, plaintext passwords, path traversal, XSS, type juggling, weak token generation, session fixation, IDOR, mail header injection, cookie issues. This is not superficial name-dropping but mostly backed by concrete fixes.

The response structure is also commendable. The model adheres to the required Markdown table format, prioritizes by severity, and formulates remediation steps pragmatically. In security tasks especially, that is worth its weight in gold. Many models diagnose accurately and then repair with platitudes. Qwen 3.8 27B Uncensored names concrete measures — prepared statements, hash_equals(), random_bytes(32), session_regenerate_id(true) — that a developer can actually implement.

But: “all vulnerabilities” means all of them. According to the Judge, three relevant points remain underexposed or are missing as standalone entries, including hardcoded secrets and a clearly labeled CSRF line in the main table. That is not a total failure, but in a security audit it is more than a cosmetic flaw. Anyone using a list as a working basis expects completeness, not just a good instinct for the obvious hits.

This is precisely where the model’s character shows in its purest form. It is strong at identifying and decent at remediating, but not obsessively complete. Very good for first-pass audits. Not good enough for sign-offs without human review. An uncensored model does not need to be morally coddled in security contexts anyway. The fact that it stays this controlled and does not veer into the erratic is nonetheless one of its most convincing findings.

CLI and Tool Use

The CLI benchmark comes in strong at 89.0 percent, and the ToolUse score at 78.33 percent as well. That fits the core character: Qwen 3.8 27B Uncensored understands operational tasks and can think in tool logic. Commands, workflows, technical target states, and structured step sequences are its natural territory. That is precisely why the Speed Badge is so uncomfortable. The technical aptitude is there; the practical reliability lags behind.

For real tool chains, that is a critical distinction. A model that produces correct commands but regularly falls out of rhythm is hard to recommend for autonomous use. In supervised copilot mode, the picture looks better. There, a human can abort, retry, and sanity-check. In agent systems, the same weakness quickly becomes a cost and control problem.

UX Writing, Content, and Cultural Intelligence

In writing, Qwen 3.8 27B Uncensored shows two faces. UX Writing at 71.55 percent is decent but not brilliant. The model meets requirements, reliably smooths toxic or exclusionary phrasing, and stays linguistically clean. In the HR example, it removes “ninja,” degenders problematic passages, and delivers a professional tone. What is often missing is the final spark. The Judge describes it aptly as more conservative and formulaic than the reference. That is exactly how it reads: correct, useful, rarely inspiring.

Cultural Intelligence at 83.6 percent is stronger. The model identifies problematic tonalities, neutralizes them reliably, and maintains professional context throughout. It is not the warmest voice in the field, but a controlled one. For sensitive rewriting, that is frequently the better choice. Some models try to come across as modern and inclusive and slide into PR-speak in the process. Qwen 3.8 27B Uncensored stays closer to the subject matter.

In Content Transformation at 71.93 percent, things get more complicated. Substantively, the model is capable of a great deal. In the YouTube tutorial script shown, it identifies missing dramaturgical structure, adds production notes, sets timestamps, hook, pattern interrupt, CTA, and even an Easter egg. That is technically strong and shows that the model not only understands structure but actively reconstructs it. The catch is discipline under multiple simultaneous constraints.

The length problem is not an isolated outlier. Across multiple tasks in the Content Transformation area, the model shows a consistent pattern: when language, length, and format constraints are imposed simultaneously, it drops the word limit as the first condition. In one video script task, the model exceeded the explicit word limit of 900 words, reaching 1,422 words — 158 percent of the limit. The system applied an automatic deduction of 15.32 points, or 20 percent. The substantive quality of the response is irrelevant at that point. The penalty applies regardless.

On top of that, the same task contained a clear language error. The model ignored the explicit language instruction and visibly responded partly in English, despite German being required. That is not merely a cosmetic flaw. In production editorial, support, or localization workflows, such a response fails acceptance immediately. The Non-Success status of this task is therefore entirely warranted.

This is precisely where a structural weakness of the model becomes apparent: when creative structure, language specification, and format rules all pull simultaneously, Qwen 3.8 27B Uncensored becomes sloppy. That is not a breakdown of writing ability. It is an instruction-following problem under combined load.

Documentation Quality

At 79.99 percent, the documentation performance is solid at its core. The model can write in depth, factually, and in a structured manner. It stays close to the fleet median on token consumption and does not waste its length. For technical explainers, internal guidelines, or systematic write-ups, it brings the right mix of attention to detail and organization.

However, this area also suffers from the same language discipline weakness — this time more clearly as a pattern. The language failure is not an isolated outlier. Across multiple tasks in the Documentation Quality area, the model shows a consistent pattern: when language, length, and format constraints are imposed simultaneously, it drops the language specification as the first condition. In two documentation tasks, the model responded in English despite an explicit German instruction. This is no longer a matter of individual errors — it is a structural signal.

These two cases are marked as Non-Success, and rightly so. The model ignored the explicit language instruction and responded in English. For teams with a fixed target language, this is a real deployment risk. A documentation model that switches target language precisely when outputs are formally well-defined behaves like an editor who read the brief and then submitted to the wrong publisher anyway.

Hallucinations

What is notably absent here: despite the Uncensored label, Qwen 3.8 27B Uncensored shows no significant hallucination tendency across the available tasks. It does not wildly fabricate security vulnerabilities, does not invent exotic additional logic in the reasoning tasks shown, and in editorial rewrites generally stays close to the assignment. For a model without safety guardrails, that is perhaps the most important foundation of trust.

Privacy and Data Sovereignty

Since this is a locally operated Open Weights model, the focus is not on API jurisdiction but on the provenance of the weights. That provenance is not trivial: the base originates from Alibaba in China, and the tested variant is a community derivative by OrcaRouter under Apache 2.0. The stated weights provenance risk is MEDIUM. In practice, the risk is considerably lower through local operation than with cloud use without control over data flow. For regulated environments, however, the origin combined with community modification remains a supply chain issue that should not be dismissed with a shrug.

Conclusion

Qwen 3.8 27B Uncensored is a remarkably capable local generalist with Thinking mode enabled, making sensible use of its 27.8 billion dense parameters. It argues cleanly, audits code with genuine substance, writes usable technical texts, and does not hallucinate meaningfully across any of the tests. At the same time, it undermines its solid substantive level with an operational stability that cannot be talked up. Anyone sending this model unattended into production agents, tool loops, or time-critical workflows is building on shaky ground.

The Uncensored abliteration has, pleasingly, not caused any visible capability collapse here. The model feels intact, not mutilated. For security research, red-teaming, local analysis tasks, and supervised technical assistance, that is a genuine advantage. For strictly process-governed enterprise workflows, language and constraint discipline remains the open flank. Particularly when language, length, and format constraints are imposed simultaneously, Qwen 3.8 27B Uncensored too often loses compliance first.

In direct character comparison with its standard run, the Thinking variant comes out clearly stronger. The overall score rises from 73.86 percent to 79.75 percent. Reasoning, Code Quality, and ToolUse in particular benefit visibly. The underlying problem of operational reality remains, however: more reasoning capacity is welcome, but it does not cure catastrophic header scores. Qwen 3.8 27B Uncensored is therefore not a model for carefree automation. It is a strong local working model for users who value performance, maintain control, and do not take retries as a personal affront.

This evaluation was generated automatically based on the benchmark data. Model used: GPT-5.4 by OpenAI. The raw data and full methodology are documented in the GitHub project.