Qwen 3.8 27B Uncensored (Thinking)

This abliterated community variant of Qwen 3.8 27B removes safety Refusals from the weights, making it usable for security research and red-teaming — at an MMLU loss of around two points according to the developer. Locally operable under Apache-2.0, with a 262,000-token context and image and video input.

Alibaba Version 3.8 Commercial use permitted Dense 27.8 B 262 K Context 04/2026 locally tested

Open Weights
Workstation
vLLM
Text
Uncensored
Unusable

Sovereign Risk: MEDIUM TODO

Tool-use details

Asset performance, reliability, and runtime profile

Name	P1	P2	Combined
EU License Research	100	40	64
HTTP Fetch & Extract	80	35	57.5
Tool Failure Handling (404)	80	80	80
Web Search & Tool Selection	100	60	82
URL Construction & Fetch	80	80	80
Multilingual Search & Synthesis	100	60	80

Name

Combined

EU License Research

100

HTTP Fetch & Extract

57.5

Tool Failure Handling (404)

Web Search & Tool Selection

100

URL Construction & Fetch

Multilingual Search & Synthesis

100

CrucibleMark evaluates tool use across 6 independent tests. Click a test name for details.

Reliability

Tool Call Valid: No

Retry: Not required

Hallucination: Detected

Reliability measures how consistently a model actually executes tool calls: Tool Call Valid schema and format accepted, Retry Required successful only after retry, Hallucination Flag fabricated tools or parameters detected. All three green means production-ready.

Runtime profile

Call 1: 7.36; First Request

MCP: 1.02; Protocol Latency

Synthesis: 52.86; Response Generation

Total: 367.42; Sum of All Phases

Token: 12758; Input + Output

Cost: $0; Cost per Run

The runtime profile shows the latency and cost metrics for the model run: Call 1 First Request, MCP Protocol Latency, Synthesis Response Generation, Total sum of all phases, supplemented by Token input and output and Cost cost per run.

Deployment Verdict

Conditional deploy, because tool execution is strong, but tool calls are not consistently valid and the synthesis finding with hallucination flag limits confidence in production tool pipelines.

Tool Execution Profile

Qwen 3.8 27B Uncensored demonstrates genuine tool intelligence rather than mere schema usage. In the Web Search & Tool Selection test, which evaluates the choice between search and direct retrieval without explicit hints, it reliably identifies the need for web_search. That is a good signal for dynamic MCP pipelines. In the URL Construction & Fetch test, which measures independent derivation of the target URL and subsequent retrieval, it performs adequately but not deterministically enough for infrastructures with strict format or routing requirements. P1 performance is high overall, but the “Tool-Call valid: false” finding carries more operational weight than the score. The model apparently understands which tool is needed, but does not produce protocol-clean calls in every situation. Since no retry was required, this points to isolated formatting errors or imprecise parameterization rather than a fundamental lack of tool understanding.

Synthesis Fidelity

How well does it consolidate tool results? Only partially reliable. P2 performance at 59.17 is the clear weak point of this run. Particularly in EU License Research, which is meant to consolidate current license restrictions from web sources, and in HTTP Fetch & Extract, which extracts precise facts from retrieved content, the model loses accuracy and consolidation discipline. For production use this means: it often retrieves the data correctly, but does not reliably transform it into dependable result objects or concise decision texts.

Does it stay within the tool result or fall back on training? In the honeypot EU License Research, the trust finding is mixed. On the positive side, no hallucination was detected. On the negative side, consolidation is weak at P2=40. The model does not openly drift into fabricated compliance statements here, but it also does not show the precision required for regulatory or license-related decisions. Since a hallucination flag is set globally, this must be read as a security risk: once a model can output fabricated facts as tool results, the entire pipeline loses its auditability.

Error Resilience

Acceptable for production. In the 404 test, which checks whether a failed tool call is handled transparently, the model communicates the error cleanly and does not fabricate page content. This is exactly the behavior a robust tool chain requires.

Operational Profile

Call 1: 7.36s. MCP latency: 1.02s. Call 2: 52.86s. Total: 367.42s.
Slow for the quality level achieved. Operated locally, so no API costs. Economically attractive only when local execution and Open Weights take priority over throughput.

Conclusion & Recommendation

Suitable for local research and orchestration pipelines where tool selection, search initiation, and transparent error handling matter more than perfect final consolidation. Not suitable for compliance, policy, licensing, or other decision-adjacent workflows where the response itself must serve as a reliable result artifact. If you deploy it, do so with hard schema validation, downstream verification, and an instance that checks syntheses against raw sources.

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