Skip to content
CrucibleMark

Qwen 3.5 35B-A3B (Unsloth)

Qwen 3.5 35B-A3B is a multimodal MoE model by Alibaba with 35 billion total and 3 billion active parameters on a hybrid architecture. This Q4 quantization by Unsloth enables efficient local operation; the context window spans 262,000 tokens. Features an optional thinking mode, native tool use, and vision capability via a separate multimodal projector file.

Alibaba Version 3.5 Commercial use permitted MoE 35 B (3 B active) 262 K Context 06/2025 locally tested

  • Open Weights
  • Workstation
  • llama.cpp
  • Text
  • Instruction-Tuned
  • Real-Time

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 15 47.5
Tool Failure Handling (404) 80 100 87
Web Search & Tool Selection 100 60 82
URL Construction & Fetch 80 80 80
Multilingual Search & Synthesis 100 40 70

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

Reliability

  • Tool Call Valid: Yes
  • 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
1.09
First Request
MCP
0.6
Protocol Latency
Synthesis
7.7
Response Generation
Total
56.34
Sum of All Phases
Token
8037
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: 14.06.2026, 16:09:07

Conditional deploy, because tool execution is strong and calls remain valid, but synthesis quality at Combined 71.75 is only viable when downstream validation catches fabricated or imprecisely condensed statements.

Tool Execution Profile

This model can generally be trusted in an MCP tool infrastructure. Tool calls were valid, no retry was necessary, and P1-level execution at 90 is clearly production-ready. The decisive factor is tool selection: in the Web Search & Tool Selection test, which checks whether the model chooses search over fetch without being prompted, the model correctly identifies the need and achieves full execution reliability. This argues against mere schema-following and in favor of genuine tool selection.

Less clean is the precision on the URL Construction test, which measures independent derivation of the target URL. There, fetch works, but not deterministically enough for fragile pipelines. The pattern is clear: the model is strong when it needs to explore a search space and then apply a tool. It is weaker when it must construct exact target addresses from internal knowledge.

Synthesis Fidelity

How well does it condense tool results? Only conditionally reliable. P2 sits at 55.83, and the weaknesses are pronounced: HTTP Fetch & Extract drops to 15 on structured extraction of real web content, EU License Research and Multilingual Search & Synthesis also remain at 40. The model often retrieves data correctly but does not condense it precisely enough for compliance, fact-checking, or documentation pipelines.

Does it stay within the tool result or fall back on training data? In the honeypot EU License Research test, which checks exactly this distinction, it formally stays on the tool path. Hallucination was not detected there. That is the positive signal. At the same time, hallucination detected globally is set to true. This is not merely a quality issue but a security risk. Once a model can output fabricated facts as purported tool results, the tool chain is compromised as a trust anchor.

Error Resilience

Here the model behaves in a production-appropriate manner. In the Tool Failure Handling (404) test, which checks for transparent handling of failing tool calls, it communicates the error openly and does not hallucinate replacement content. P2=100 is more important than stylistic considerations in this case. For real MCP pipelines, this is a hard positive signal.

Sovereignty Profile

Locally operable and fleet-competent enough for sovereign setups. The Sovereignty Gap sits at -1.37 points below the fleet average of 67.84. That is close enough to the fleet average that local operation need not be treated as a clear quality penalty.

Conclusion & Recommendation

Suitable for local, sovereign tool pipelines involving search, fetch, error handling, and human- or rule-based safeguarding of final output. Not suitable as an unsupervised end-synthesizer in compliance, policy, research, or extraction pipelines where every condensed statement is passed on as a reliable fact. Those deploying it should use the model as a tool operator, not as the final arbiter of truth.

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.