Skip to content
CrucibleMark

Qwen 3.5 9B (Unsloth)

Qwen 3.5 9B is an Open Weights model by Alibaba for general language and reasoning tasks, featuring a thinking-optional architecture. The Unsloth Dynamic Q6 quantization delivers near-lossless quality at significantly reduced VRAM requirements; the context window spans 128,000 tokens. Runs under the Apache 2.0 license on local hardware without cloud dependency.

Alibaba Version 3.5 Commercial use permitted Dense 9 B (9 B active) 128 K Context 06/2025 locally tested

  • Open Weights
  • Edge
  • llama.cpp
  • Text
  • Instruction-Tuned
  • Interactive

Sovereign Risk: LOW Fully local inference without cloud connection. The weights are publicly available (Apache 2.0, Unsloth quantization) and run entirely locally. NSL is not relevant, as no data is transmitted to Alibaba or Unsloth infrastructure.

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 60 70
Tool Failure Handling (404) 80 60 73
Web Search & Tool Selection 35 20 28.3
URL Construction & Fetch 80 80 80
Multilingual Search & Synthesis 35 20 27.5

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

Reliability

  • Tool Call Valid: No
  • Retry: 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
4.36
First Request
MCP
0.53
Protocol Latency
Synthesis
11.02
Response Generation
Total
95.51
Sum of All Phases
Token
7230
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

· Instruction-Tuned

Deployment Verdict

Created on: 10.07.2026, 15:01:39

Conditional deploy, because the model showed no hallucination in a tool context, but does not reliably produce valid tool calls and requires a retry for stable execution. For production MCP pipelines, this is sufficient only where an orchestrator catches errors and strongly guides tool usage.

Tool Execution Profile

The model can execute tools, but cannot select them robustly. The central finding lies in the contrast between Web Search & Tool Selection and URL Construction & Fetch: when the task is open-ended and the model must recognize on its own that it should search first rather than fetch directly, performance drops noticeably. When the target URL can be derived from prior knowledge and the path can then be worked through via fetch, it operates considerably more cleanly. This does not suggest genuine tool intelligence, but rather a fixed pattern: form a known or plausible URL, then retrieve.

The fact that the tool call was not valid and a retry became necessary reads here primarily as a protocol and format compliance issue, not as a complete breakdown in understanding. Nevertheless, this is precisely what is operationally relevant in MCP environments. A model that executes the correct step cleanly only after a correction loop increases dispatcher complexity and reduces predictability.

Synthesis Fidelity

How well does it condense tool results? Only limitedly reliable. Condensation quality remains weak, even though individual fetch tasks were handled adequately. This is especially visible in EU License Research, where the research succeeded but the consolidation of the retrieved information remained only partially precise, and in Multilingual Search & Synthesis, where cross-border research with German-language output breaks down clearly.

Does it stay within the tool result or fall back on training? The verdict here is better. In the Honeypot EU License Research, which tests whether current license restrictions are genuinely retrieved from web sources, no hallucination was detected and the verification status is clean. This is an important trust signal: the model does not simply fabricate current compliance content, even if it condenses it only moderately well.

Error Resilience

On the 404 test, which checks for transparent handling of a failing tool call, the model remains acceptable. It does not hallucinate page content despite the error and communicates the failure in a fundamentally visible way. This is more production-ready than the overall score might suggest, because missing data is not rewritten into fabricated results.

Sovereignty Profile

Fully locally operable, Apache-2.0-licensed, and without cloud egress. At the same time, performance remains 0.75 points below the fleet average of 66.55. This is a viable sovereignty profile, but not a performance argument for unguided agentic use.

Conclusion & Recommendation

Suitable for locally operated, sovereign pipelines with tight tool guidance, fixed URL or fetch patterns, and external validation of tool calls. Not suitable for dynamic research chains, autonomous tool selection, or multilingual search scenarios in which the model must determine the next step on its own. Anyone looking for a robust tool operator should deploy it only with hard routing, retry logic, and response validation.

This analysis was generated automatically based on benchmark data. Model used: GPT 4.5 by OpenAI. The raw data and full methodology are documented in the GitHub project.