Models & providers

The in-app chat runs on whichever model you pick. Dalea supports nine providers: one that's hosted by us out of the box, and eight that you bring your own API key for.

The Dalea default, no setup

The Dalea provider runs on operator-managed EU infrastructure and is enabled for every workspace from day one. You don't need to register an account anywhere, you don't store a key. Use it for general-purpose chat, prototyping, and any workspace that doesn't want to manage external billing. A user who never opens the picker is on Dalea AI, the label for the platform default engine.

DeepSeek V4 Flash

The default engine (`deepseek-v4-flash-0731`). Strongest reasoner on the hosted lineup, with cached input. Text-only on the wire; images are handled by the vision delegate. 256K ctx

Qwen3.6 35B A3B (MoE)

Open-weight mixture-of-experts, natively multimodal. Also the delegate that answers image questions for the text-only default. 256K ctx

GLM-5.2

Hosted GLM endpoint, text-only. 256K ctx

Two Anthropic models are also keyless. Claude Sonnet 5 and Claude Haiku 4.5 run on the operator's own Anthropic key and appear in the same "Dalea AI" list in the picker and in the AI Providers settings card, no vault key needed. The server advertises them only on a deployment that has an operator Anthropic key configured. They meter against your daily allowance rather than an Anthropic bill of your own; Sonnet 5 consumes it about 3x faster than Haiku.

Bring-your-own-key providers

For the eight key-required providers, each user adds their own API key from Settings → AI Providers. Keys are personal, they live in your user vault, not the workspace's, so you control them and you can revoke at any time. Keys are encrypted at rest and never logged. The request goes straight from our servers to the provider.

Anthropic

Opus 4.8, Opus 4.7 and Sonnet 5 use adaptive thinking: the effort tier ships as an output-config setting and the legacy thinking budget is rejected. Haiku 4.5 stays on the classic thinking budget, which its effort tiers map onto. Every Claude entry accepts images.

Claude Opus 4.8

Frontier model, the default Opus. 1M ctx

Claude Opus 4.7

Previous frontier Opus, still supported. 1M ctx

Claude Sonnet 5

Balanced speed and quality, and the widest tier range. Also available keyless. 1M ctx

Claude Haiku 4.5

Fastest Claude with thinking. Also available keyless. 200K ctx

OpenAI

GPT-5.x family. Every entry reasons and accepts images.

GPT-5.6 Terra

Newest 5.x flagship. 1M ctx

GPT-5.6 Luna

Sibling 5.6 flagship. 1M ctx

GPT-5.5 Pro

Pro tier. `low` effort is not accepted, minimum is `medium`. 1M ctx

GPT-5.5

Previous flagship. 1M ctx

GPT-5.4

Older flagship; still supported for parity. 1M ctx

GPT-5.4 Mini

Cheaper / faster. Caps at `high`, no `xhigh`. 1M ctx

Google

Gemini 2.5+ and the open-weight Gemma 4 family, all multimodal. Gemini models expose effort tiers; Gemma 4 has no effort knob, and Dalea does not enable its thinking mode.

Gemini 3.1 Pro (preview)

Frontier Gemini, preview channel. 1M ctx

Gemini 3.1 Pro (preview, tool-tuned)

Same model tuned to prefer user-supplied tools, recommended for tool-heavy chats. 1M ctx

Gemini 3 Flash (preview)

Cheaper preview. 1M ctx

Gemini 2.5 Pro

Stable flagship. 1M ctx

Gemini 2.5 Flash

Faster, cheaper sibling. 1M ctx

Gemini 2.5 Flash-Lite

Smallest Gemini 2.5 tier. 1M ctx

Gemma 4 31B / 26B A4B

Open-weight, Apache 2.0. Natively multimodal. 256K ctx

Mistral

Mistral splits into adjustable-reasoning, native-reasoning, and non-reasoning families. All entries are vision-capable. The unified Mistral 3.x / 4 generation (Medium 3.5, Small 4, Large 3) ships at 256K context; Magistral medium/small stay on the 128K reasoning-tuned base.

Magistral Medium / Small

Native reasoners, always think, no effort knob. 128K ctx

Mistral Medium 3.5

Reasoning toggle. 256K ctx

Mistral Small 4

Same toggle, smaller and cheaper. 256K ctx

Mistral Large 3

Non-reasoning multimodal MoE. 256K ctx

DeepSeek

V4 family, text-only: DeepSeek's production API does not yet accept images.

DeepSeek V4 Pro

Flagship V4. 1M ctx

DeepSeek V4 Flash

Cheaper V4 variant with the same reasoning tiers. 1M ctx

Z.ai

Zhipu's GLM models on the international Z.ai API. Both default to thinking on and are text-only at the API layer.

GLM-5.2

Flagship GLM. 1M ctx

GLM-5.1

Previous generation. 200K ctx

Moonshot AI

Kimi K-series on api.moonshot.ai. Both are native reasoners with no user effort knob, and both accept images.

Kimi K3

Flagship MoE, natively multimodal. 1M ctx

Kimi K2.6

Smaller sibling. 256K ctx

Qwen (Alibaba)

Commercial Qwen over the international DashScope OpenAI-compatible endpoint. These are wired as standard chat models: thinking is off and there is no effort knob.

Qwen3-Max

Flagship agent model, text-only. 1M ctx

Qwen-Plus

Cheaper general model, text-only. 1M ctx

Qwen3-VL Plus

Multimodal Qwen. 256K ctx

Reasoning effort

For models that expose a controllable reasoning knob, the picker shows the effort tiers the provider actually accepts. They vary:

Provider	Effort tiers
Anthropic (Opus 4.8 / Opus 4.7, adaptive)	low / medium / high / xhigh
Anthropic (Sonnet 5, adaptive)	low / medium / high / xhigh / max
Anthropic (Haiku 4.5, thinking budget)	low / medium / high
OpenAI (GPT-5.6 Terra / Luna, GPT-5.5, GPT-5.4)	low / medium / high / xhigh
OpenAI (GPT-5.5 Pro)	medium / high / xhigh
OpenAI (GPT-5.4 Mini)	low / medium / high
Google (Gemini 2.5 + 3.1 Pro)	low / medium / high
Google (Gemini 3 Flash)	minimal / low / medium / high
Mistral (Medium 3.5 / Small 4)	none / high
DeepSeek (V4 Pro / Flash)	high / max
Z.ai (GLM-5.2 / GLM-5.1)	high / max
Dalea (DeepSeek V4 Flash)	low / high / max
Dalea (GLM-5.2)	high / max

Native reasoners have no dial at all: Magistral, Kimi K3 / K2.6 and the hosted Qwen3.6 always think, and the picker shows the reasoning badge without tiers. Gemma 4 is listed without tiers too, but Dalea does not enable its thinking mode. Higher effort buys deeper analysis at higher latency and cost.

Vision

The composer's image affordance follows the model's actual route to images. Models with native vision accept image parts on the wire. Text-only models on the Dalea-hosted provider still take images: the server routes them to a vision-capable delegate through its view_image tool. Only a text-only bring-your-own-key model (DeepSeek V4, Z.ai GLM, Qwen3-Max, Qwen-Plus) has no route at all, and there the composer disables the paperclip and ignores image drops. The server rejects image attachments for those models too, defence in depth.

A turn carries at most 6 images totalling 2 MB of decoded image data. Both are fixed platform constants, not operator-tunable settings. Compress screenshots before attaching if you hit the limit.

Picking and switching

Each user has a default model (Settings → AI Providers), and the chat header lets you pick a different one when you start a chat. A conversation is locked to the model of its first turn. Once a chat is grounded, the picker still shows the catalogue but selections become no-ops, and the server rejects a turn that names a different model with CONVERSATION_MODEL_LOCKED. To switch engines, open a new chat. The default in Settings applies to new chats only.

What's next