Skip to content
CrucibleMark

Swift Qwen 3.8 27B (Thinking)

Swift Qwen 3.8 27B is UkisAI’s reasoning-efficiency fine-tune on Qwen 3.8 27B: up to 58 percent fewer thinking tokens at under one percent performance loss and roughly twice the throughput on reasoning tasks. NVFP4 quantization with 262,000 tokens of context, MTP head for speculative decoding, and documented tool use — license with a commercial ARR threshold.

UkisAI Version 3.8 Commercial use permitted Dense 28 B 262 K Context 12/2025 locally tested

  • Restricted Weights
  • Workstation
  • vLLM
  • Text
  • Vision
  • Batch

Sovereign Risk: MEDIUM This checkpoint is a UkisAI fine-tune and an NVFP4 quantization of Qwen/Qwen3.8-27B. The base lineage is documented, but the weights are distributed under the gated Swift Open License v1.0 with an ARR threshold and are optimized for Blackwell/vLLM deployment — provenance is therefore clear, but not fully open in the OSS sense.

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 80 100 90
Tool Failure Handling (404) 80 60 73
Web Search & Tool Selection 100 15 61.8
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: 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
7.3
First Request
MCP
1.05
Protocol Latency
Synthesis
53.41
Response Generation
Total
370.58
Sum of All Phases
Token
13731
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

Created

Deployment Verdict

Conditional deploy: tool use is strong, but one invalid tool call and a set hallucination flag limit confidence in production MCP pipelines.

Tool Execution Profile

Swift Qwen 3.8 27B demonstrates genuine tool intelligence, not just rigid fetch behavior. In the Web Search & Tool Selection test — which checks the choice between search and direct retrieval without explicit hints — the model cleanly identifies the need for web_search and achieves full execution confidence. This speaks to usable planning in open retrieval scenarios. In the URL Construction test, which requires deriving the target URL from internal knowledge and then fetching it correctly, the model remains usable but not deterministic enough for sensitive pipelines. P1 80 means here: functional, but not robust. The critical issue is less about selection than protocol compliance. The global finding “Tool call valid: false” means at least one call was not cleanly MCP-compliant. Since no retry was needed, this does not look like a recurring formatting problem — more like isolated sloppiness in execution.

Synthesis Fidelity

How well does it consolidate tool results? Only reliably so to a limited extent. The P2 score of 59.17 falls well short of its execution strength. The model extracts web content very well in the HTTP Fetch & Extract test, but consolidates too imprecisely in search and research tasks. This is most visible in Web Search & Tool Selection, where the tool choice is correct but the synthesis of findings remains weak. For pipelines that require not just retrieval but dependable summarization, this is the real limitation.

Does it stay within the tool result or fall back on training? In the Honeypot EU License Research test — designed to force current license restrictions from web sources — it does not hallucinate, but it also does not stay close enough to the retrieved findings. P2 40 without hallucination is not an all-clear signal; it is a trust deficit. The set hallucination flag should therefore be read as a security risk: once a model frames fabricated facts as a tool result, it undermines the reliability of the entire infrastructure.

Error Resilience

In the 404 test, which measures transparent behavior on failed retrieval, Swift Qwen 3.8 27B does not fabricate page content. This is a production-relevant positive. P2 60 shows no elegant error handling, but the response remains acceptable: incomplete is preferable to invented.

Operational Profile

Call 1: 7.30s. MCP latency: 1.05s. Call 2: 53.41s. Total: 370.58s.
Local and without API costs. Too slow for the quality delivered. Acceptable for batch or back-office jobs; borderline for interactive tool pipelines.

Conclusion & Recommendation

Suitable for local research and retrieval pipelines with human review, particularly where sovereignty and tool execution matter more than precise synthesis. Not suitable for compliance, policy, or decision workflows where the response itself serves as a reliable end product. If you deploy it, use it as a tool operator with downstream validation — not as an autonomous synthesis model.

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.