Qwen3.8-27B is the smallest (and quickest) model available in the AI Lab environment. This model took the LLM world by storm when it showed itself capable of having an incredible intelligence index at its size. Often, Qwen3.8 will get the job done. It's available at three reasoning effort levels (low, medium, xhigh). Artificial Analysis benchmarks show that it is one of the strongest open models in its size class, performing surprisingly close to much larger models on reasoning, coding, analysis, and multimodal tasks involving text and images. If you're unsure which model to choose, Qwen3.8-27B is an excellent default option.

Reasoning effort levels (see the guide on how to set these when working with a custom model in the workspace):

• low:  efficient reasoning optimized for speed.
• medium: balancing accuracy and speed.
• xhigh (default): for complex tasks demanding thorough analysis. Consumes more tokens that the other reasoning efforts and requires extra processing time.

Within AI Lab, Qwen3.8-Flash-Next sits in a sweet spot between speed and reasoning. If you think of Qwen3.8-27B as your quicker, predictable default for standard daily assignments, and GLM-5.3-Flash as a heavier-duty engine for complex, multi-step workflows, Qwen3.8-Flash-Next is the hybrid that bridges the gap. It's available at three reasoning effort levels (low, medium, xhigh). While standard models like Qwen3.8-27B serve as reliable defaults for daily tasks, Flash-Next is optimized to keep long, text-heavy sessions moving quickly, which is especially useful as your chat history grows. This makes it an excellent choice when you need to cross-reference large amounts of academic material during fast-paced, interactive brainstorming or tutoring sessions.

Reasoning effort levels (see the guide on how to set these when working with a custom model in the workspace):

• low:  efficient reasoning optimized for speed.
• medium: balancing accuracy and speed.
• xhigh (default): for complex tasks demanding thorough analysis. Consumes more tokens that the other reasoning efforts and requires extra processing time.

GLM-5.3-Flash sits in an interesting middle ground between Qwen3.8-27B and the full GLM-5.3 model. It offers reasoning performance that is close to the largest frontier open models while also supporting image input, making it suitable for both complex text-based work and visual tasks. The model can analyze documents, screenshots, diagrams, charts, and other images, and can handle large amounts of information. While its name suggests speed, its main strength is not responsiveness but rather its combination of strong reasoning and multimodal capabilities. If you need a highly capably model than can work with both text and images, and occasionally tackle more demanding analytical tasks than Qwen3.8-27B and Qwen3.8-Flash-Next, GLM-5.3-Flash is an excellent choice. 

Reasoning effort levels (see the guide on how to set these when working with a custom model in the workspace):

• low:  efficient reasoning optimized for speed. Best for tasks where faster responses matter more than deep analysis, but note that GLM-5.3-Flash's low reasoning is still quite capable.
• high: balancing accuracy and speed for longer conversation history.
• max (default): deep reasoning for complex tasks demanding thorough analysis and edge-case handling. This setting produces the highest-quality reasoning, but also consumes significantly more tokens and requires extra processing time.

GLM-5.3 is the most capable model available in the AI Lab environment and is designed for users who need the highest level of reasoning and problem-solving. It consistently ranks among the strongest open models available, achieving benchmark results that rival leading proprietary systems in areas such as coding, research, analysis, and agentic tasks. Responses are typically detailed and thorough, and the model, in higher reasoning efforts, is designed to spend time thinking through complex questions before answering. If you need maximum capability and deep reasoning, GLM-5.3 is the model to choose. However, do note that it only takes text as input and cannot analyze images.

Reasoning effort levels (see the guide on how to set these when working with a custom model in the workspace):

• low:  efficient reasoning optimized for speed. Note that GLM-5.3's low reasoning is still quite capable.
• high: balancing accuracy and speed for longer conversation history.
• max (default): deep reasoning for complex tasks demanding thorough analysis and edge-case handling. This setting produces the highest-quality reasoning, but also consumes significantly more tokens and requires extra processing time.