Hosted Models

GenAI Studio provides open-weight models for general chat, reasoning, coding, tool use, and image understanding. The inventory below was verified against the GenAI Studio models API on August 31, 2026. Model availability may change as deployments are added, upgraded, or retired.

Use the model ID exactly as shown when sending an API request. To retrieve the models available to your account at any time, query the authenticated https://genai.rcac.purdue.edu/api/models endpoint or open the All section of the model selector in the web interface.

This page lists the base models hosted by RCAC. Custom models that you create or that another user shares with you may also appear in the selector and API response; their availability and behavior depend on their Workspace configuration.

Start with a recommended model

The four vLLM-backed models are tagged Recommended in GenAI Studio. Use gpt-oss:120b for general reasoning, qwen3.6:27b for coding and agentic workflows, or gemma4:26b-a4b and llama4:latest when your prompt includes images.

Recommended Models

These models run on vLLM with automatic tool selection enabled. The context limit is the deployed limit, not the larger native limit that some upstream model cards advertise.

Model ID	Inputs	Good for	Deployed context
gpt-oss:120b	Text	General-purpose and high-effort reasoning, coding, and agentic tasks	65,536 tokens
llama4:latest	Text, image	Multimodal reasoning, image understanding, coding, and general chat	16,384 tokens
gemma4:26b-a4b	Text, image	Multimodal reasoning, coding, structured output, and tool-based workflows	65,536 tokens
qwen3.6:27b	Text, image	Reasoning, software development, tool use, and agentic workflows	65,536 tokens

All four vLLM deployments support automatic tool calling and model-specific reasoning output. For structured output through the API, vLLM models use response_format; Ollama models use format. See Structured Output for examples.

Ollama Models

The following models run on Ollama with a maximum configured context length of 65,536 tokens. A model with a smaller native context window may have a lower effective limit; setting a larger server limit does not extend what the model was trained to use reliably.

The model IDs follow the tags published in the Ollama model library. Use the complete ID shown below; IDs that include a quantization suffix select that specific variant.

Model family	Available model IDs	Inputs	Good for
Code Llama	codellama:latest	Text	Code generation, explanation, and completion
DeepSeek-R1	deepseek-r1:1.5b deepseek-r1:7b deepseek-r1:14b deepseek-r1:32b	Text	Reasoning, mathematics, and step-by-step problem solving; larger variants generally provide stronger results at higher latency
Devstral Small 2	devstral-small-2:latest	Text	Repository-level coding and software-engineering agents
Gemma 3	gemma3:1b	Text	Lightweight chat, summarization, and simple instructions
Gemma 3	gemma3:12b gemma3:27b	Text, image	General chat, reasoning, multilingual tasks, and image understanding
GPT-OSS	gpt-oss:latest	Text	Reasoning, coding, and tool-oriented tasks in the Ollama deployment
Llama 3.1	llama3.1:latest llama3.1:70b-instruct-q4_K_M	Text	General instruction following, summarization, and long-form work; use the 70B variant when response quality is more important than latency
Llama 3.2	llama3.2:latest	Text	Lightweight general chat and instruction following
Llama 3.3	llama3.3:70b	Text	High-quality multilingual chat and instruction following
LLaVA	llava:latest	Text, image	Visual question answering and general image understanding
MedGemma	medgemma:27b	Text, image	Research and experimentation involving medical text and image comprehension
Mistral	mistral:latest	Text	Fast general chat, summarization, and instruction following
Phi-4	phi4:latest	Text	Reasoning, mathematics, and compact general-purpose workflows
Qwen 2.5	qwen2.5:72b	Text	General multilingual work, reasoning, and coding
Qwen 3	qwen3:1.7b qwen3:4b qwen3:8b qwen3:14b qwen3:30b qwen3:32b	Text	General chat, reasoning, multilingual work, and coding; smaller variants favor speed while larger variants favor quality
Qwen 3 Coder	qwen3-coder:latest	Text	Agentic coding, repository analysis, and software development
Qwen 3 VL	qwen3-vl:32b	Text, image	Document and image understanding, visual reasoning, and multimodal workflows
QwQ	qwq:latest qwq:32b-fp16	Text	Deliberate reasoning, mathematics, and complex problem solving

Ollama supports tool calling only when the selected model and template support it. Do not infer tool-call compatibility from a model being visible in GenAI Studio. Test the intended workflow before depending on it, or choose one of the recommended vLLM models with an explicitly configured tool parser. See Tool Calling for the supported UI and API flows.

Medical model limitations

medgemma:27b is provided for research and experimentation. Model output is not medical advice and must not be used as a substitute for review by a qualified professional. Do not submit protected health information or other regulated data to GenAI Studio.

Document Embedding Model

GenAI Studio uses EmbeddingGemma 300M internally to create text embeddings for Knowledge Base search and retrieval. Embeddings allow the service to find document passages that are semantically related to a prompt before sending the relevant passages to the selected chat model.

EmbeddingGemma is not a generative chat model and does not appear in the model selector. RCAC does not expose it as a callable model or through a standalone embeddings API. GenAI Studio applies it automatically when processing documents uploaded through the Knowledge Base interface.

Practical Limitations

The context window includes the system prompt, conversation history, uploaded or retrieved document excerpts, tool definitions and results, image tokens, and the model's response. Leave room for the response instead of filling the entire deployed limit with input.

Model IDs ending in :latest are aliases rather than fixed versions. RCAC may update the model behind an alias without changing its ID. For reproducible work, record the model ID, request settings, date, prompt, and output; do not rely on a :latest alias remaining unchanged.

GenAI Studio is a shared service. Response time varies with model size, prompt length, output length, and current demand. A larger model is not always better for a particular task. Evaluate models with representative prompts before selecting one for a research or production workflow, and see API Rate Limits before running concurrent workloads.

Back to Purdue GenAI Studio