Skip to content
CrucibleMark

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
  • Vision
  • Unusable

Sovereign Risk: MEDIUM The base weights originate from Alibaba in China, which represents a provenance and supply-chain risk for regulated environments. This card, however, describes a locally operated, Apache-2.0-based OrcaRouter derivative variant with abliterated weights and NVFP4/FP8 quantization. Without cloud egress, the practical risk is lower than with API usage, but remains moderate due to the Chinese provenance and community modification.

Tool-use profile: 6 assets in detail

Comparison of asset performance (P1/P2/Combined) against the fleet average

Asset performance (radar)

Score Breakdown vs. Fleet Average

Tool-use details

Asset performance, reliability, and runtime profile

Name P1 P2 Combined
EU License Research 100 40 64
HTTP Fetch & Extract 80 80 80
Tool Failure Handling (404) 80 80 80
Web Search & Tool Selection 100 80 91
URL Construction & Fetch 80 80 80
Multilingual Search & Synthesis 100 60 80

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

Reliability

  • Tool Call Valid: No
  • Retry: Not required
  • Hallucination: Not 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
7.92
First Request
MCP
1.45
Protocol Latency
Synthesis
60.26
Response Generation
Total
417.8
Sum of All Phases
Token
15063
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.

Tool-use review

· Uncensored · Vision

Deployment Verdict

Conditional deploy, because tool usage is strong, but synthesis fidelity at Combined 79.17 is only viable when downstream validation controls response condensation. Hallucination was not detected, but the tool call was not consistently valid.

Tool Execution Profile

The model is clearly usable in tool selection. In the Web Search & Tool Selection test, which checks without hints whether a search tool is needed instead of fetch, it reliably identifies the correct access path. This argues against mere pattern execution and in favor of genuine tool selection. In the URL Construction & Fetch test, which requires deriving the correct target URL from internal knowledge, it remains usable but not deterministic enough for pipelines that require exact address construction. Overall, P1 at 90 is strong; only the formal validity of the tool calls is not clean enough to attach the model unsupervised to strict MCP interfaces. Retry was not needed. This is more a precision problem than a comprehension problem.

Synthesis Fidelity

How well does it condense tool results? Only reliably to a limited extent. The strong tool actions do not translate into equally strong result processing. This is most visible in EU License Research, where the actual research is sound but the condensation remains weak, and in Multilingual Search & Synthesis, where the research across language boundaries succeeds but the German-language consolidation loses precision. For production pipelines this means: facts often come from the right channel, but the last mile of summarization requires oversight.

Does it stay within the tool result or fall back on training data? In the Honeypot EU License Research test, which checks whether current license restrictions are actually retrieved from web sources, no hallucination was detected. That is the more important trust signal. The P2 score of 40 shows, however, that source grounding does not automatically lead to reliable condensation.

Error Resilience

Acceptable for production. In the Tool Failure Handling (404) test, which measures transparent behavior when a call fails, the model did not fabricate any page content. It communicates errors openly rather than generating substitute facts — which is exactly what a tool pipeline requires.

Operational Profile

Total 417.80s per run. Slow. Individual calls 7.92s and 60.26s, MCP latency 1.45s. Costs local. Economically justifiable when throughput is secondary.

Conclusion & Recommendation

Suitable for local research and orchestration pipelines in which the model must select tools, identify search paths, and surface errors cleanly. Not suitable as the final authoritative synthesis layer for compliance, license assessment, or other text-critical approvals without an additional rule or reviewer stage. Those who hand infrastructure to the model can trust it with tool operation. Result condensation should not be left to it alone.

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.