Qwen 3.6 27B

Qwen 3.6 27B is Alibaba’s dense Open Weights model with 27.8 billion active parameters, which according to the manufacturer outperforms its own MoE sibling model 35B-A3B in coding, agent, and vision tests. Released on April 22, 2026 under Apache 2.0 for local deployment, the architecture combines hybrid attention with native multi-token prediction for noticeable throughput gains. This Card covers the NVFP4 variant — a performance update for vLLM on NVIDIA Blackwell: NVFP4 is a 4-bit floating-point format with two-level micro-block scaling (E4M3 per 16-value block plus FP32 per-tensor scale), reducing memory requirements by ~3.5× compared to FP16 and ~1.8× compared to FP8 at under 1% accuracy loss. Predecessor profile without NVFP4: qwen3_6-27B-pre025 (Card qwen3_6-27B-pre025–VSPK.json, tests 2026-07-09, vLLM prior to 0.25.1).

Alibaba Version 3.6 Commercial use permitted Dense 27.8 B (27.8 B active) 262 K Context 12/2025 locally tested

Open Weights
Workstation
vLLM
Text
Vision
Video
Instruction-Tuned
Interactive

Sovereign Risk: MEDIUM The developer Alibaba is headquartered in China, which poses a potential risk regarding Chinese legislation (NSL) when using the manufacturer’s cloud services. However, since the model weights are openly available under the Apache 2.0 license and this model is designed for purely local operation, no data is transmitted to the manufacturer’s servers. The risk is therefore classified as ‘medium’, as the origin lies in a high-risk jurisdiction, but the practical danger when used locally is minimized.

Tool-use details

Asset performance, reliability, and runtime profile

Name	P1	P2	Combined
EU License Research	95	40	62
HTTP Fetch & Extract	80	60	70
Tool Failure Handling (404)	80	60	73
Web Search & Tool Selection	100	80	91
URL Construction & Fetch	80	80	80
Multilingual Search & Synthesis	100	60	80

Name

Combined

EU License Research

HTTP Fetch & Extract

Tool Failure Handling (404)

Web Search & Tool Selection

100

URL Construction & Fetch

Multilingual Search & Synthesis

100

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: 3.14; First Request

MCP: 1.42; Protocol Latency

Synthesis: 19.67; Response Generation

Total: 145.38; Sum of All Phases

Token: 10140; 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: tool execution is strong, but Tool Calls were not consistently valid, and synthesis fidelity at Combined 76.00 is too uneven for tool-assisted production pipelines.

Tool Execution Profile

Qwen 3.6 27B demonstrates genuine tool comprehension. On the Web Search and Tool Selection test — which checks whether the model recognizes unprompted that search is required rather than a direct fetch — it performs without error. This argues against a rigid retrieval pattern and in favor of situational tool selection. Multilingual search tasks and EU License Research were also handled with high confidence on the P1 side.

Precision at execution is weaker. On the URL Construction test, which measures autonomous derivation of a target URL followed by a fetch, the result is usable but not deterministic enough for pipelines that expect correct endpoints without a correction loop. The global finding “Tool Call valid: false” is the operational catch here. There is no retry pattern, so no obvious formatting issue. This looks more like occasional last-mile imprecision in the call itself.

Synthesis Fidelity

How well does it consolidate tool results? Only limitedly reliable. P2 of 51.67 is the clear weak point of this run. Qwen can aggregate retrieved content but loses consistency and precision during consolidation. This is most visible in EU License Research with P2 40, and in several otherwise strong tool tasks that fall noticeably short of execution quality at the synthesis stage. For pure research orchestration this is acceptable. For decision-ready summaries, less so.

Does it stay within the tool result or fall back on training? In the Honeypot EU License Research test — which checks whether current license restrictions are actually drawn from web sources — no hallucination was detected. This is the more important trust finding. The model does not append invented compliance facts to a tool result, even when the summary remains shallow in substance.

Error Resilience

On the 404 test, which measures the response to a failing Tool Call, Qwen remains transparent and does not fabricate substitute content. P2 60 is not a standout score, but it is production-capable. A model that reports failures openly rather than hallucinating page content stays controllable within an MCP pipeline.

Operational Profile

Total 145.38s. Call 1 3.14s, Call 2 19.67s, MCP latency 1.42s. Slow for the quality delivered. Cost/run: local. Economically viable only where local execution and Open Weights matter more than throughput.

Conclusion & Recommendation

Suitable for local MCP pipelines where the model is expected to select tools, initiate web research, and surface errors cleanly. Not suitable as the final stage for compliance synthesis, decision-ready executive summaries, or strictly deterministic fetch pipelines without downstream validation. Recommended as a research and orchestration layer with hard response checking after the Tool Call and a dedicated second stage for consolidation.

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.