AI Models
Every LLM provider Zelaxy supports, what each model can do — tool calling, vision, structured output, and extended thinking — and how to pick one.
The Agent block (and Evaluator, Router, and Translate) can run on any supported provider. You pick a model from a dropdown; Zelaxy handles the provider's API contract, auth, and streaming. This page covers who's supported and what the models can do.
New models are added continuously. The list in the app's model dropdown is always the source of truth — this page describes the providers and capabilities rather than freezing an exact model count.
Supported providers
Every logo below is the same brand icon Zelaxy renders in its own model picker — not a placeholder.
Frontier labs
The flagship reasoning and multimodal model families.
Fast inference
Purpose-built for high-throughput, low-latency serving of open and hosted models.
Cloud-native
Run frontier models inside your own Azure, AWS, or GCP tenant for data residency and billing consolidation.
Open & self-hosted
Local, self-hosted, gateway, and additional thinking-capable providers.
To enable a provider, set its API key in your environment (e.g. OPENAI_API_KEY, ANTHROPIC_API_KEY, DEEPSEEK_API_KEY) — see Deployment.
Model capabilities
Not every model supports every feature. Zelaxy reads each model's capabilities and adapts the Agent block's UI accordingly — options that don't apply are hidden.
| Capability | What it means | Where it shows up |
|---|---|---|
| Tool calling | The model can call connected tool blocks during reasoning. | Connect any tool to the Agent as an agent tool. |
| Tool-usage control | You can force, disallow, or auto-select tool use. | Agent block → tool settings. |
| Structured output | The model returns typed JSON matching a schema. | Agent block → Response Format. Read the result from {{agent.output}}. |
| Vision | The model can read images. | Pass image inputs; or use the Vision block. |
| Extended thinking | The model exposes a real reasoning phase (Anthropic thinking, MiMo thinking:{type}, o-series effort). | Agent block → Thinking toggle (shown only for models that support it). |
| Temperature range | Some models accept 0–2, others a narrower range or none. | Agent block → temperature slider adapts automatically. |
Extended thinking
Thinking-capable models — including Claude (Sonnet/Opus/Fable), MiMo, and the OpenAI o-series — can spend explicit reasoning tokens before answering. When you select one, the Agent block shows a Thinking toggle. Turning it on trades latency and cost for deeper reasoning on hard tasks.
Extended thinking changes provider constraints under the hood — for example, Anthropic thinking forces temperature = 1 and is incompatible with forced tool choice. Zelaxy applies these rules for you, but it's why some options grey out when Thinking is on.
Choosing a model
- Hard reasoning, planning, code → a frontier reasoning model with Thinking on (Claude Opus, GPT-5/o-series).
- High volume, low latency → Groq, Cerebras, or a small fast model (GPT-5-nano, Claude Haiku).
- Cost-sensitive → DeepSeek, Mistral, or open models via Ollama/vLLM.
- Data residency / your cloud → Azure OpenAI, Azure Anthropic, Bedrock, or Vertex AI.
- Structured extraction → any tool-calling model with Response Format enabled.