Qwen3.6-Plus: Towards real world agents

Model openness & business strategy

• Qwen3.6-Plus is closed-weight and hosted-only; parameter count is undisclosed.
• Many see this as a pivot from the reputation built on open-weight releases, interpreting prior open models as “marketing loss leaders” rather than altruism.
• Others note that Plus/Max/Omni variants were always closed, so nothing fundamentally changed; smaller open-weight variants are promised again.
• Some users say they mainly care that small–medium models remain open; others want large open models on ethical grounds, given training on public data.

Benchmark comparisons & “misleading” claims

• Strong concern that Qwen compares against Claude Opus 4.5 (not 4.6) and Gemini 3.0 (not 3.1), and omits certain OpenAI coding benchmarks.
• Critics call this deceptive “previous-gen” marketing; defenders argue the timelines are tight and 4.5 is still a familiar, meaningful reference point.
• Several note Qwen is still behind last-gen Opus in many metrics, undercutting “SOTA” claims.

Market for non-SOTA models

• Debate over whether a real market exists for “almost-SOTA but cheaper”:
  • One side: “everyone wants the best.”
  • Other side: cost dominates in production, large batch data tasks, and sub-agent orchestration; “good enough” cheap models are valuable.
• Examples include using cheaper models for automated workflows, sub-agents, and customer-facing tasks where perfect quality isn’t required.

Privacy, trust & geopolitics

• Some distrust Alibaba-hosted models and prefer US providers; others trust none except local inference.
• Others prefer Chinese providers over US ones, arguing their own government is the more immediate threat and that US surveillance practices are also aggressive.
• Broader geopolitical tensions (US vs China, allied countries caught between) strongly influence provider choice.

Real-world performance & agent behavior

• Reports of Qwen3.6-Plus hallucinating more than some competitors and ignoring explicit “planning-only” instructions, getting stuck in loops, especially with tools.
• Others report excellent agentic benchmark scores and say open Qwen 3.5 variants already perform similarly well in agent tasks.
• Some users feel personality, adherence to instructions, and token efficiency matter more than raw benchmark scores.

Open ecosystem & future

• Several expect Chinese labs to keep open-sourcing mid/large models because they lack strong direct sales channels and rely on open releases for visibility and distribution.
• There’s enthusiasm for large-context open models and for Qwen’s generous free quotas via CLI and promotions, even among skeptical users.