Qwen 3.6 35B-A3B (Unsloth) (Thinking)

Qwen 3.6 35B-A3B is Alibaba’s MoE model with 35 billion total and approximately 3 billion active parameters per token, released on April 22, 2026 under Apache 2.0 with open weights for local deployment. The hybrid attention architecture combines classic attention with a linear variant; Multi-Token Prediction noticeably accelerates generation.

Alibaba Version 3.6 Commercial use permitted MoE 35 B (3 B active) 262 K Context 12/2025 locally tested

Open Weights
Workstation
vLLM
Text
Vision
Video
Instruction-Tuned
Interactive

Sovereign Risk: MEDIUM The model originates from the Qwen team (Alibaba), based in China. The classification of risk as ‘medium’ rather than ‘high’ reflects that this is an open-source model under the permissive Apache 2.0 license, which can be run entirely locally without any cloud connection to Alibaba servers. In purely local operation, NSL relevance is virtually eliminated; a theoretical residual risk due to the Chinese developer jurisdiction remains for the purposes of the provenance assessment.

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	95	80	88.3
URL Construction & Fetch	80	60	70
Multilingual Search & Synthesis	95	60	77.5

Name

Combined

EU License Research

100

HTTP Fetch & Extract

Tool Failure Handling (404)

Web Search & Tool Selection

88.3

URL Construction & Fetch

Multilingual Search & Synthesis

77.5

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: 9; First Request

MCP: 1.19; Protocol Latency

Synthesis: 27.5; Response Generation

Total: 226.16; Sum of All Phases

Token: 26544; 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: The model is fundamentally usable for tool execution, but the combination of only moderate synthesis fidelity, an invalid tool call, and a Combined Score of 76.62 is not sufficient for unattended production pipelines.

Tool Execution Profile

Qwen 3.6 35B-A3B demonstrates genuine tool intelligence, but not consistently clean protocol discipline. On the Web Search & Tool Selection test — which checks whether the model searches first rather than fetching directly, without being prompted — it identifies the correct tool class very reliably. This argues against rigid pattern behavior. On the URL Construction test, which measures the autonomous derivation of a target URL followed by a fetch, it remains usable but not deterministic enough for fragile pipelines. The core finding is therefore clear: it selects tools mostly sensibly, but does not always produce valid MCP-compliant calls. The absence of any retry needed points more toward inconsistent call form than to a comprehension problem with the task flow.

Synthesis Fidelity

How well does it consolidate tool results? Only conditionally reliable. The P2 score of 56.67 shows that Qwen often merges retrieved content correctly, but does not maintain precision, prioritization, and consolidation stably enough. This is also visible in EU License Research: correct tool usage, but weak final consolidation. By contrast, HTTP Fetch & Extract and Tool Failure Handling (404) are noticeably cleaner in their summaries.

Does it stay within the tool result or fall back on training data? In the honeypot EU License Research — which tests exactly this breach of trust — it does not hallucinate. That is the key safety finding. The weak P2 score of 40 is therefore more a consolidation problem than a trust problem. For compliance-adjacent workflows this is better than fabricated currency, but not yet a free pass.

Error Resilience

On the 404 test — which measures transparent handling of failed tool calls rather than fabricated fallback content — the model responds in a production-appropriate manner. It does not hallucinate page content despite the error and communicates the failure with sufficient transparency. For real MCP pipelines this is a strong signal, because a tool failure may reduce response quality, but must not destroy the factual basis.

Operational Profile

Total 226.16s per run. Call 1: 9.00s. MCP latency: 1.19s. Call 2: 27.50s. Operated locally, therefore no API costs. Clearly slow for the performance shown.

Conclusion & Recommendation

Suitable for locally operated research, retrieval, and assistant pipelines with human review or downstream validation. Particularly useful where tool failures must be handled cleanly and local weights matter more than maximum response precision. Not the right choice for fully automated compliance, policy, or decision workflows in which every tool call must be formally valid and every synthesis must stay closely aligned with the tool output.

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.