Qwen3.6-27B: Dense VLM

Overview

Qwen3.6-27B is a 27B-parameter dense model from Qwen (Alibaba Group). It’s a causal language model with a built-in vision encoder, so the same checkpoint can handle text-only chat and vision-language reasoning. The model can generate a hidden reasoning trace (“thinking”) and then a final answer, which helps on multi-step coding and analysis. It is useful for coding tasks, document analysis, and multi-turn debugging, within the roughly 25,000 input tokens a request takes on Wiro.

What you can build

• Code review and refactoring assistants that keep project rules across threads of up to about 25K tokens
• Agentic coding helpers that plan changes, then generate patches and tests
• Repository Q&A over file excerpts, logs, and build output of up to about 25K tokens
• Technical writing tools for API docs, release notes, and migration guides
• Document analysis for specs, contracts, and incident postmortems of up to about 25K tokens
• Vision-language support flows like screenshot-to-explanation (when your deployment supports vision)

Inputs

• A required user message written as plain text. This is the task, question, or coding request.
• An optional “thinking” toggle that controls whether the model shows its reasoning trace before the answer.
• An optional system instruction written as plain text. Put role, style, and safety rules here.
• Optional user and session identifiers as short strings. Use them to keep chat history across turns.
• Optional randomness controls that shape how deterministic or varied the text is.
• Optional sampling controls that limit generation to high-probability tokens for steadier outputs.
• Optional repetition controls that reduce looping and repeated phrases.
• Optional length controls that bias toward shorter or longer completions.
• Optional minimum and maximum output length limits measured in tokens.
• Optional stop phrases, provided as a semicolon-separated list. Generation stops at the first match.
• An optional numeric seed to make sampling more repeatable between runs.
• An optional switch to force deterministic decoding (no sampling).
• An optional quantization switch to reduce memory use, with some quality trade-offs.

Outputs

A text response that contains the model’s completion for your request.

If thinking is enabled, the response may include a reasoning block before the final answer (often wrapped in think-style markers), followed by a clean user-facing result. For coding tasks, the output often includes code blocks, step lists, and structured diffs when you ask for them.

Recommended settings

• Precise coding and refactors: keep randomness low (0.6) and sample from a high-probability set (top-p 0.95, top-k 20). Keep repetition penalty at 1.0.
• General reasoning: increase randomness (1.0) while keeping top-p 0.95 and top-k 20. Keep repetition penalty at 1.0.
• Direct answers without thinking: use moderate randomness (0.7) and a tighter probability mass (top-p 0.8, top-k 20). Keep repetition penalty at 1.0.

Limitations

• The model can hallucinate facts, filenames, test results, or tool outputs. Verify anything important.
• Vision and video understanding depend on the serving stack. Some deployments expose only text chat.
• On Wiro, a single request takes in about 25,000 input tokens, counting the system message, chat history, tool definitions, and your prompt together. The checkpoint's 262,144-token window is not available here.
• Longer input is shortened automatically, without an error: the oldest history messages are dropped first, then the end of your latest message is cut. Put your question at the start of long prompts.
• Responses are capped at 16,384 output tokens, and thinking tokens count toward that cap.
• Long reasoning traces can crowd out space for the final answer if you set large output limits.
• Low-quality inputs reduce accuracy. This includes noisy logs, inconsistent formatting, and pasted code with missing context.
• If you feed scanned images, blurry screenshots, or dense UI captures, vision answers may be wrong.

Safety & compliance

• Don’t use the model to generate instructions for wrongdoing, malware, or exploitation.
• Treat generated code as untrusted. Review for security issues, data exfiltration, and unsafe shell commands.
• Don’t rely on outputs for medical, legal, or financial decisions without qualified review.
• Follow the model’s Apache 2.0 license terms when you redistribute weights or derivatives.