You can spawn AI apps now

spawnpoint now runs AI models on its own servers, and your app can use them. They are open-source models, which cost a fraction of the closed models from the big labs to run, and they run on spawnpoint's machines, next to your app. Try it: Runway is a choose-your-own-adventure through San Francisco startup life, written as you play by an open model on spawnpoint: Demo Day in 48 hours, a Series A pitch with a golden retriever in the room, a Friday outage with the one engineer who knows the deploy out of signal. Pick one of its three choices, or type what you do instead. The line under each scene shows how long the model took to write it.

Your agent decides when an app gets this, the way it decides when an app gets a database. An app it writes from scratch that needs a model uses an open model from the start. An app you already have that uses OpenAI or Anthropic is different, because you may have reasons to stay with them. Your agent asks you once: switch to an open model on spawnpoint, or keep what you have. If you keep it, nothing changes, and your agent does not ask again for that app. Each app that switches gets its own key to the models, created when the app goes live and deleted when the app is deleted. Right now there is no extra charge for the models. If that changes, we will say so first.

The models available today, and what each is for:

Model	Good for
gpt-oss-120b	Conversation, writing, summarizing, answering questions. The default.
gpt-oss-20b	The same, faster, for apps that answer a lot of questions.
glm-5.2	Apps written for Anthropic's format, kept as they are.
gemma-4-31b, glm-5.3-flash	Describing, reading, or comparing images.
qwen3.8-27b	Conversation in many languages.
bge-m3	Finding the closest match in a pile of your own text, which is how a help search works.
laya	Quick typed decisions: which category, how urgent, yes or no. More on this below.

Making images, speech, and video are not on the list yet. An app that does those keeps using the company it already uses for them, and your agent will tell you so.

We measured the models from an app on spawnpoint, the way your app would use them, asking each the same question five times with the answer capped at 300 tokens, a token being a word or part of one. The middle run of the five:

Model	First word after	Tokens a second
gpt-oss-20b, quick mode	0.50 s	556
gpt-oss-20b	0.46 s	213
gpt-oss-120b, quick mode	0.48 s	214
gpt-oss-120b	0.61 s	60
qwen3.8-27b	0.22 s	138
gemma-4-31b	0.42 s	200
glm-5.3-flash	0.24 s	55
glm-5.2	1.14 s	82

The gpt-oss models think before they answer, and quick mode shortens that thinking, which is why Runway uses gpt-oss-20b in quick mode and has the first words of a scene on screen in under a second. Turning a sentence into a search vector with bge-m3 took 53 milliseconds in the middle run of ten.

Decision models are the part most people have not met. A chat model writes an answer one word at a time, which is right when you want text. A decision model reads your text once and gives back a typed answer: a category from a list you supply, a score on a scale you supply, or a yes or no, each with a probability. It never writes anything, so it answers in a fraction of a second, is billed only for what it reads, at 2 cents per million tokens, and your app gets a value it can act on without reading a sentence. The kind of job it suits: a support inbox that files each message to the right team and flags the angry ones, a comment box that holds back spam, a form that sorts requests by how urgent they sound, a feed that tags each post with a topic.

Runway uses both. When you type your own move instead of picking a choice, laya reads the scene and your move and answers three questions in one pass: is the move possible here, is it likely to work, and is it safe, risky, or career-ending. That verdict goes to gpt-oss-20b, which writes the scene around it, so a reckless move can end the story, and the line above the scene shows what laya decided and how long it took. A decision model is the right pick when the categories are clear and you want speed, a probability, and a cost you can ignore, and a chat model is the right pick when the call needs judgment. Your agent knows both, and picks by what the app does.