--- title: "Model Profiles" description: "Control which LLM model and settings your assistant uses for each job — conversations, memory, background tasks — and optimize token costs with profiles and call-site overrides." canonical_url: "https://www.vellum.ai/docs/key-concepts/model-profiles" md_url: "https://www.vellum.ai/md/docs/key-concepts/model-profiles" related: - "/docs/key-concepts" - "/docs/key-concepts/channels" - "/docs/key-concepts/memory-and-context" - "/docs/key-concepts/oauth-integrations" - "/docs/key-concepts/scheduling" - "/docs/key-concepts/skills-and-tools" - "/docs/key-concepts/the-workspace" - "/docs/key-concepts/web-search" --- # Model Profiles Control which LLM your assistant uses for each job — conversations, memory work, scheduled tasks — and override it per call-site when you need to. ## Overview Model profiles control which AI model your assistant uses and how it behaves. You set one profile as the workspace-wide default, and your assistant applies it to everything — conversations, background memory work, scheduled tasks, and more. You can override it per conversation or per action type when you need different behavior. ## Built-in profiles Every workspace starts with three built-in profiles. You can edit or duplicate them, but the defaults can't be deleted. | Profile | Model | Best for | | -------------- | ------------ | ---------------------------------------------------------------------- | | Quality | Claude Opus | Deep research, complex reasoning, high-stakes tasks | | Balanced | MiniMax M3 | Everyday use — capable across the board at reasonable cost | | Cost Optimized | Claude Haiku | Simple, short, or structural tasks where speed matters more than depth | **Balanced is active by default.** Every call your assistant makes — conversation replies, memory filing, title generation — runs with the Balanced profile unless you change it. ## Switching profiles Open **Settings → Inference Profiles** and select a different active profile from the dropdown. The change applies workspace-wide immediately — no restart needed. You can also create custom profiles from this screen if you want to use a different provider or model not covered by the built-ins. ## Per-conversation override Use the profile picker in the chat header to pin a different profile to a single conversation. It only affects that conversation — your workspace default stays untouched. This is the easiest way to run one session on a stronger or lighter model without changing anything globally. ## Our recommendation Keep **Balanced** as your active profile — it covers everyday use well. Then use Action Overrides to selectively upgrade the actions that benefit most from a stronger model, and step down only for tasks where the output is purely structural. A good rule of thumb: anything that shows up directly in your conversation or drives a decision should stay on at least Sonnet. Tasks like generating a title, formatting a notification, or suggesting conversation starters are good candidates for a lighter model — they're short, easy to verify, and quality differences are barely noticeable. **Tip:** use the per-conversation profile picker for one-off heavy tasks. Switch to Quality for a deep research session, then leave your workspace default untouched. That way you only pay for Opus when you explicitly reach for it. ## Action overrides Open **Settings → Inference Profiles → Action Overrides** to assign a specific profile to individual actions. Each action has a toggle — when off it uses your active profile, when on you pick a profile just for that action. You can search by name and reset everything back to defaults at any time. | Action | What it does | Recommended | | ---------------------------- | ----------------------------------------------------------------- | -------------- | | Agent loop | | | | Main agent | The primary conversation agent that handles your messages | Quality | | Subagent spawn | Spawns a subagent to handle a delegated subtask | Quality | | Heartbeat agent | Runs background tasks and proactive checks on a schedule | Balanced | | Filing agent | Files memories and updates the knowledge base after conversations | Balanced | | Analyze conversation | Analyzes conversation content for summaries and insights | Balanced | | Call agent | Handles voice call conversations | Quality | | Memory | | | | Memory · Extraction | Pulls facts and preferences out of conversations and stores them | Balanced | | Memory · Consolidation | Merges and deduplicates your memory store over time | Balanced | | Memory · Retrieval | Searches memory to surface relevant context during conversations | Balanced | | Narrative refinement | Refines and polishes stored narrative memory entries | Balanced | | Pattern scan | Scans conversation history to detect behavioral patterns | Balanced | | Conversation summarization | Summarizes long conversation threads for memory and context | Balanced | | Conversation starters | Generates suggested openers for new conversations | Cost Optimized | | Workspace | | | | Conversation title | Generates a title for each conversation | Cost Optimized | | Commit message generator | Writes git commit messages from staged changes | Balanced | | UI | | | | Identity intro | Generates your assistant's introductory message on first launch | Balanced | | Empty-state greeting | Generates the greeting shown on an empty conversation | Cost Optimized | | Notifications | | | | Notification decision | Decides whether to surface a proactive notification to you | Balanced | | Preference extraction | Learns your communication preferences from how you interact | Balanced | | Voice | | | | Guardian question copy | Generates spoken prompts during guardian verification flows | Balanced | | Watch commentary | Produces live commentary delivered via Apple Watch | Balanced | | Watch summary | Generates brief summaries surfaced on Apple Watch | Cost Optimized | | Utility | | | | Interaction classifier | Classifies the type of each inbound message to route it correctly | Cost Optimized | | Style analyzer | Analyzes your writing style to help your assistant match it | Balanced | | Invite instruction generator | Generates onboarding instructions for new assistant invites | Balanced | | Skill category inference | Automatically categorizes installed skills | Cost Optimized | | Skills | | | | Meet · Consent monitor | Monitors meeting consent during Google Meet sessions | Balanced | | Meet · Chat opportunity | Identifies moments to send a helpful message during meetings | Balanced |