Models
Models and providers

Models

4. Models#

RayuCode v3 reads its model catalog live from the provider. With a Rayu credential, that catalog is whatever your account or key may use: labels, descriptions, context windows, thinking support and vision support all come from the server, so a new model appears without updating rayu.

The authoritative list for your account is in the picker and at https://rayucode.com/models. Model ids and availability change; treat any id in these docs as an example.

Choosing a model#

In a session, type:

text
/model

The picker:

  • filters by typing, matching the model id,
  • moves with ↑ / ↓, selects with Enter, closes with Esc,
  • shows the context window and capability badges for each entry.

Ctrl+N toggles a favorite, Ctrl+O sets the session default, and Shift+Tab cycles favorites without opening the picker.

Set a model without the picker:

text
/model deepseek-v4.1-flash-cloud

or at launch:

bash
rayu --model deepseek-v4.1-flash-cloud
rayu model list                # every id you can pass to --model
rayu model list --json
rayu model list --verbose      # include the provider selection summary

Your choice persists across sessions and is written to [provider] in ~/.rayu/config.toml.

Model ids vs labels#

Use the id anywhere Rayu takes a model: --model, /model, the "model" field of a Rayu API request, and [provider] default_model. The label in the picker is display only.

Example ids from a real catalog (yours may differ):

IdNotes
DeepSeekdeepseek-v4.1-flash-cloudfast, 1M context, thinking
DeepSeekdeepseek-v4.1-flash-1017-cloudpinned build of the same family
longcat-2.5long context
kimi-k3.1-procode-oriented
glm-5.3-flash-cloudthinking
claude-sonnet-5.5thinking, vision
claude-opus-5.5strongest Claude route
claude-haiku-5.5fastest Claude route
gpt-6-astrathinking, vision
grok-4.7thinking
MiniMaxminimax-m3.1-flash-cloudthinking
Qwenqwen4-max-previewlong context

Aliases like sonnet, opus and haiku resolve to the Claude routes when the catalog carries them.

Reasoning effort#

Reasoning-capable models expose an effort level. Higher effort spends more reasoning tokens; it is not free.

text
/effort none
/effort minimal
/effort low
/effort medium
/effort high
/effort xhigh
/effort max
/effort swarm

swarm runs the turn in swarm mode (see Subagents & the swarm). The current level is shown in /info, and Alt+Left / Alt+Right cycles it.

Fast mode is separate: /fast toggles a provider priority tier when the model and account support one.

Context windows#

/context prints the session's assembled system prompt sections, the effective tool list, the active messages, and how much of the context window is used. /compact summarizes older messages to free room; /clear starts a new conversation.

The header shows a live context meter. When it approaches the model's limit, Rayu compacts automatically according to [compaction].

Vision and file attachments#

Paste an image into the prompt, or reference one, and any vision-capable model sees it. The picker marks which models accept images. Sending an image to a text-only model fails with a clear error.

Images that tools return (screenshots, rendered diagrams, PDFs) also render inline in the transcript.

Per-role models#

Rayu uses more than one model: the main conversation, spawned swarm workers, the automatic reviewer, the automatic judge, and memory extraction.

text
/agents

/agents shows the configured roles and lets you set a model per role. Global defaults live in config.toml:

toml
[agents]
swarm_model = "deepseek-v4.1-flash-cloud"   # spawned workers
swarm_effort = "low"
memory_model = "deepseek-v4.1-flash-cloud"
autoselect = true                            # or false to always inherit

[autoreview]
enabled = false

[autojudge]
enabled = false

/subagent-model shows or changes the worker policy for this session, and /autoreview / /autojudge toggle the end-of-turn review and judge passes.

Next#

Next: Configuration →