Model guide
Qwen 3.8 27B Uncensored overview #
This imageat deployment packages Qwen 3.8 27B Uncensored as a hosted chat API: no model download, inference server, or GPU capacity planning is required. Applications send familiar system, user, and assistant messages, while imageat handles execution, credit settlement, and the final OpenAI-style response.
The model is useful when a request needs more room than a typical short-context assistant—large specifications, multi-file code excerpts, research notes, or extended conversations. Thinking mode is a per-request choice, so one integration can support both quick chat and more deliberate analysis.
What you can build #
- Add long-context assistance to an imageat workflow without hosting a separate LLM service
- Review large product specifications, logs, or code excerpts submitted by your backend
- Generate structured creative briefs before sending work into image and video pipelines
- Build support agents that preserve role-based conversation context across several turns
- Turn research notes into implementation plans, checklists, documentation, or reports
- Prototype prompts in the playground and reuse the selected parameters through the API
Outputs #
The model returns a single assistant response as UTF-8 text in an OpenAI-compatible chat completion response. When thinking mode is enabled, the response may include a dedicated reasoning block before the final answer.
choices[0].message.content
Read the generated assistant text from this field.
Model inputs #
Control the response style, length, repeatability, and conversation identity through the imageat Chat Completions API.
messages
Your main instructions, questions, and conversation history as role-based text messages.
system message
An optional message that sets the assistant’s role, tone, and behavioral rules.
enable_thinking
Enables deeper thinking-style output or disables it for shorter, direct responses.
user & session_id
Optional identifiers used to keep separate users and conversation sessions.
temperature, top_p & top_k
Randomness and probability-sampling controls for generation.
repetition_penalty
Reduces repeated phrases and generation loops.
length_penalty
Nudges the model toward shorter or longer completions.
min_tokens & max_tokens
Sets optional minimum and maximum output-token limits.
stop
A string or array of phrases that stops generation at the first match.
seed
An optional integer seed for more repeatable sampled outputs.
quantization
Trades a small amount of quality for more efficient inference when enabled.
do_sample
Switches between randomized sampling and more deterministic generation.
Recommended settings #
Use these presets as a starting point based on the official Qwen 3.8 sampling guidance.
Reasoning and coding
Enable thinking, then use temperature: 1.0
, top_p: 0.95
, top_k: 20
, and repetition_penalty: 1.0
.
Direct answers and chat
Disable thinking, then use temperature: 0.7
, top_p: 0.80
, top_k: 20
, and repetition_penalty: 1.0
.
Limitations #
- Reduced refusal behavior is not a guarantee that every prompt will be answered or that every answer is appropriate.
- Messages and generated output share the 262,144-token context budget; reserve enough space for the completion.
- Large context capacity does not guarantee perfect retrieval—important instructions should be explicit and well structured.
- Thinking mode can materially increase runtime, output length, and the final imageat credit tier.
- Sampling and quantization settings can change factual consistency, formatting, and reproducibility.
- The imageat endpoint documented here accepts text messages; it does not expose the checkpoint's possible vision inputs.
Safety and compliance #
This model may be more willing to answer sensitive requests than strongly aligned chat models. You must still follow imageat platform rules and all applicable laws. Do not use it for instructions that enable wrongdoing, violence, self-harm, or illegal access.
For user-facing products, add your own moderation, logging, rate limits, abuse prevention, and human review for high-risk use cases.