Models
Typillar generates code with a model that runs on your side, and it’s Cloudflare-native: every model in the catalog is reached through your own Cloudflare connection. There are no keys to paste and no other provider to configure — connecting Cloudflare is all it takes. No inference runs on Typillar’s servers; see Zero inference on our servers.
Your Cloudflare connection is the model source
Section titled “Your Cloudflare connection is the model source”Once you’ve connected Cloudflare, the agent has its model. Tokens are generated on Cloudflare’s network and billed through your Cloudflare account. Model access is part of the baseline Cloudflare grant, so there’s nothing extra to enable — type an idea and the agent can plan and build.
The catalog
Section titled “The catalog”The build loop is agentic: the agent drives its tools (write a file, read a file, finish) through function calls, so every model in the catalog is one that supports function calling. There are twenty, in three groups — what separates them is what it takes on your Cloudflare account to reach them. Of the twenty, thirteen read images too — hand one a screenshot or a mockup and it builds from what it sees — and each of those rows is marked Sees images. — and seven of those read PDFs as well — a spec, a brief, a paper — marked Reads PDFs. The split is the wire’s doing, not the models’: Claude and GPT-5.x ride API endpoints whose schemas carry a file part, while the chat-completions endpoint the rest use refuses one, so Gemma, Kimi, DeepSeek, Gemini and Qwen see pixels but never pages. Grok is the odd row: xAI’s own API reads images, but the only image part Cloudflare’s gateway will carry is the one xAI’s endpoint rejects, so no screenshot reaches Grok on this path — hand a mockup to one of the marked rows instead.
Included — Workers AI, any plan. These run on Workers AI under your account and are billed as ordinary Workers AI usage, against the same free allowance every Cloudflare account gets:
| Model | ID | Notes |
|---|---|---|
| GPT-OSS 120B | @cf/openai/gpt-oss-120b |
The default. 128K context |
| Qwen 3.8 27B | @cf/qwen/qwen3.8-27b |
Sees images. 256K context |
| Nemotron 3 120B | @cf/nvidia/nemotron-3-120b-a12b |
256K context |
| GLM-4.7 Flash | @cf/zai-org/glm-4.7-flash |
128K context |
Workers Paid — Workers AI behind a plan gate. Also Workers AI, also billed as ordinary Workers AI usage, but Cloudflare puts these four behind standard Workers Free billing: reaching them takes either the Workers Paid plan or prepaid AI Gateway credits.
| Model | ID | Notes |
|---|---|---|
| GLM-5.2 | @cf/zai-org/glm-5.2 |
256K context |
| Kimi K2.7 Code | @cf/moonshotai/kimi-k2.7-code |
Sees images. 256K context |
| DeepSeek V4 Pro | @cf/deepseek-ai/deepseek-v4-pro-0813 |
Sees images. Beta at Cloudflare. 1M context |
| DeepSeek V4 Flash | @cf/deepseek-ai/deepseek-v4-flash-0731 |
1M context |
On a free Cloudflare account with no credits these are refused, and Typillar tells you so instead of quietly building with something else. The selector marks them Paid plan. See What Cloudflare’s free plan gives you.
Credits — frontier models on your Cloudflare billing. Cloudflare’s catalog also carries closed frontier models, reachable through the same connection:
| Model | ID | Notes |
|---|---|---|
| Claude Opus 5 | anthropic/claude-opus-5 |
Sees images. Reads PDFs. 1M context |
| Claude Fable 5 | anthropic/claude-fable-5 |
Sees images. Reads PDFs. 1M context |
| Claude Sonnet 5 | anthropic/claude-sonnet-5 |
Sees images. Reads PDFs. 1M context |
| GPT-5.6 Sol | openai/gpt-5.6-sol |
Sees images. Reads PDFs. 1M context |
| GPT-5.6 Terra | openai/gpt-5.6-terra |
Sees images. Reads PDFs. 1M context |
| Gemini 3.7 Flash | google/gemini-3.7-flash |
Sees images. 1M context |
| MiniMax M3 | minimax/m3 |
1M context |
| Kimi K3 | moonshotai/kimi-k3 |
Sees images. 1M context |
| Qwen 3.8 Max | alibaba/qwen3.8-max |
Sees images. 1M context |
| Grok 4.6 | xai/grok-4.6 |
500K context |
| Claude Haiku 4.5 | anthropic/claude-haiku-4.5 |
Sees images. Reads PDFs. 200K context |
| GPT-5.6 Luna | openai/gpt-5.6-luna |
Sees images. Reads PDFs. 1M context |
Selecting one routes the build through Cloudflare’s AI Gateway unified billing: you pay each provider’s own per-token rate on credits you load in your Cloudflare dashboard, with no markup from Typillar. Still your account, still your bill, still no inference on our servers. Load the credits before you pick one — without them the provider call is refused, and Typillar surfaces that error rather than silently downgrading you to something cheaper.
Choosing one
Section titled “Choosing one”Pick a project default — or a one-off override for a single build — from the
model selector next to the composer. Your choice governs builds. The Plan
agent always reasons with the default model, @cf/openai/gpt-oss-120b.
An unknown or retired choice never dead-ends a build: it falls back to the default rather than failing.
Why the model runs on your account
Section titled “Why the model runs on your account”- Cost transparency. You see and pay for inference directly, on your own account — no markup, no proxy, no surprise line item.
- One connection, one bill. Your app, its data, its hosting, and its model all live on the same Cloudflare account — nothing else to wire up.
- A clean trust boundary. Because Typillar never runs the model, your prompts and generated code don’t pass through our inference servers.
Related
Section titled “Related”- Your models — why inference runs on your side.
- Your data & secrets — how secrets are protected.