Qwen / Qwen3.8-27B
Qwen3.8-27B
Qwen3.8-27B is a 27B-parameter chat model from Qwen with optional thinking traces and 262k native context for long-form coding, analysis, and agent workflows.
Overview
Qwen3.8-27B is a 27B-parameter causal language model from Qwen, released as open weights under Apache 2.0. It generates a final answer, and it can also produce an explicit reasoning trace before the answer when “thinking” is enabled. It’s built for long-horizon work like repo-level coding, multi-step problem solving, and long-document analysis, where short-context chat models often lose key details. The base checkpoint is a vision-language model with a vision encoder, but this Wiro listing focuses on text chat controls.
What you can build
- Long-context code review and refactoring plans for large files and multi-module projects
- Agent-style task plans that break a goal into steps and track intermediate decisions
- Deep technical Q&A that needs careful multi-step reasoning
- Long-document summarization and cross-referencing across sections
- Structured drafting for specs, design docs, and technical reports
Inputs
- Your main user prompt as plain text.
- Optional system instructions as plain text that set rules and tone for the session.
- An optional toggle that shows the model’s reasoning trace, or hides it and returns only the final answer.
- Optional user identifier text to help keep history tied to a person across runs.
- Optional session identifier text to separate multiple chats for the same user.
- A randomness control (numeric) that shifts outputs from deterministic to more varied.
- A nucleus sampling threshold between 0 and 1 to limit token selection to the most likely mass.
- A top-k cutoff (integer) to sample only from the k most likely next tokens.
- A repetition control (numeric) to reduce loops and repeated phrases.
- A length bias control (numeric) that nudges outputs shorter or longer.
- A maximum output length in tokens.
- A minimum output length in tokens, with an option to disable the minimum.
- Optional stop strings, provided as a semicolon-separated list, that end generation when matched.
- An optional numeric seed for repeatable sampling.
- An optional low-memory mode toggle that runs the model with quantized weights.
- An optional toggle that switches between deterministic decoding and sampling-based decoding.
Outputs
The model returns generated text for the assistant message.
If thinking is enabled, the output may include a reasoning section before the final answer. Many runtimes format this as a distinct block (often wrapped in tags like …) followed by the user-facing response.
Recommended settings
- For deep reasoning with thinking enabled: set randomness to 1.0, nucleus threshold to 0.95, and top-k to 20.
- For direct instruction answers with thinking disabled: set randomness to 0.7, nucleus threshold to 0.80, and top-k to 20.
Limitations
- The model’s native context window is 262,144 tokens. Some stacks can extend it up to 1,000,000 tokens, but quality can vary at extreme lengths.
- Very long prompts can raise memory use sharply and may force smaller batches or shorter outputs.
- If you enable thinking traces, the model can produce long reasoning blocks that crowd out the final answer budget.
- Low-quality input reduces reliability. Expect worse results with messy logs, missing context, or contradictory requirements.
- Quantized runs can reduce output quality, especially on delicate reasoning or exact formatting.
- Stop strings can cut off answers mid-sentence if they match common text.
Safety & compliance
Follow Qwen’s Usage Policy when you generate or share content.
Do not use the model to generate or promote illegal activity, child sexual abuse material, privacy violations, violence or terrorism, fraud, or instructions to bypass safeguards. Avoid prompt-injection or “jailbreak” attempts, and don’t request sensitive personal data about private individuals.
API quick start
Run Qwen3.8-27B with a single API call.
{
"prompt": "Explain the Second Law of Thermodynamics …",
"user_id": "...",
"session_id": "...",
"enableThinking": "true"
}