Choosing the company that runs your model

Since September 20, 2026, the tryingopen composer lets you choose which company runs your model. An open model's weights are public, so many companies serve the same model on their own GPUs. The picker works through OpenRouter, which can send your message to any company that serves the model.

Which companies the picker lists

Next to the model button sits a small provider button that lists the companies able to run your model. As of October 2026, our allowlist holds 41 companies, among them CoreWeave, Baseten, Modal, Groq, Cerebras, Together, DeepInfra, Fireworks, Amazon Bedrock and Nebius. Each model lists only the ones that serve it. GLM 5.3 Flash has 21.

A company is offered for a model only if a request pinned to it with OpenRouter's zero-retention flag succeeded, with one exception covered below. OpenRouter's docs say that with the flag set, a request "will only be routed to endpoints that have a Zero Data Retention policy." We sent each model to each company OpenRouter lists for it, with the flag on, and dropped any pair that kept failing.

What a pick does

A pick is strict. Your message goes to that company only, with fallbacks switched off, and the flag stays on. In the terms of OpenRouter's provider routing, a CoreWeave pick looks like this:

"provider": {
  "only": ["coreweave"],
  "allow_fallbacks": false,
  "zdr": true
}

If CoreWeave can't take the message, you get an error and the message goes nowhere else. The same holds if a company ever loses its zero-retention status: the request fails.

A reply from a company you picked names it, as "via CoreWeave" beside the model's name. "Get the code for this reply" turns it into a curl command for OpenRouter's API with the same model and provider settings, to run with your own OpenRouter key. What each reply tells you about how it was made walks through the rest.

Until you pick, the popover's top row is in use. For most models it reads "Let OpenRouter choose" and picks a zero-retention provider for each message. On "Provider may log" models it is the maker's or host's own API (see below). Gemma 4 31B, Gemma 4 26B and GPT-OSS 120B go from that row to Amazon Bedrock through ngrok's AI gateway, marked "Zero retention" under the composer.

A pick carries over when you switch to another model that company serves. When it doesn't serve the new model, the picker moves to CoreWeave if CoreWeave serves it and it isn't a "Provider may log" model, and to the top row otherwise. CoreWeave is our choice for that spot, and the popover tags its row "Default".

The line under the composer names the company in use: "Open model · Run by CoreWeave · Zero retention" with CoreWeave picked, and "Provider picked per message" in the middle with "Let OpenRouter choose".

A page for every company

Each of the 41 companies has its own page on tryingopen, such as /coreweave, /baseten and /modal. A provider page is the ordinary chat with that company already picked. The /coreweave page is titled "Trying CoreWeave: open models, served by CoreWeave", opens on GLM 5.3 Flash, and shows the same CoreWeave caption under the composer.

Since October 13, 2026, the page address carries your setup: the provider sits in the path and the model in the query. Copying the address shares that combination. Sharing a chat, a model setup or a preset as a link explains what goes into a link.

Three sites with one company each

We also launched three brand sites with the picker on September 20, 2026: tryingcoreweave.com, tryingbaseten.com and tryingmodal.com. Each locks the provider, so every message goes to that one company. The banner on tryingcoreweave.com reads "Every message goes to CoreWeave with a zero data retention requirement."

CoreWeave's and Modal's sites start on GLM 5.3 Flash. Baseten's starts on DeepSeek V4.1 Flash.

All three are our sites, and each footer says so. On tryingcoreweave.com it reads "Built on Tryingopen by Offline Base. Not affiliated with CoreWeave." The other two name Baseten and Modal. Our privacy policy names all three.

What "Provider may log" means

Since October 7, 2026, some models go by default through ngrok's AI gateway straight to the API of their maker or host. Those companies may keep what you send under their own terms, so the picker labels these models "Provider may log". The privacy page puts it this way: "Other models are reached through an AI gateway run by ngrok instead, unless you pick a provider for them."

As of October 2026, nine models carry the label: the eight on the ngrok route, and Mistral Large 4, which goes through OpenRouter and has no zero-retention host yet. For five, picking a provider sends the message through OpenRouter with zero retention. The other four have no such route today.

With no pick, the line under the composer for GLM 5.3 Flash reads "Run by Z.ai · Provider may log". The first row of the provider popover, "Z.ai, through ngrok", says "The default. It may keep your message; pick a provider below for zero retention." For the four without a route, hovering over "Provider may log" explains that the model "has no zero-retention host yet, so its provider may keep your message."

The exception mentioned earlier is Alibaba, the one company listed for Qwen3.8 Flash: a message sent there also goes without the zero-retention flag, so the label stays. MiMo V2.6 Flash, MiMo V2.6 Pro and Mistral Large 4 have no provider button.

When the edge model hands off to the cloud

The "Provider may log" default matters most for GLM 5.3 Flash, the cloud model behind tryingopen's default pick. That pick, LFM2.5-VL 3B ("Runs on an edge device"), answers quick questions on our own hardware and sends harder ones to a cloud model: GLM 5.3 Flash unless you pick another. With no provider picked, those cloud turns go to Z.ai through the gateway, under the terms above.

Since October 8, 2026, a provider picker sits beside the cloud-fallback picker and chooses the company for the fallback model. Pick CoreWeave there and the cloud turns go to CoreWeave through OpenRouter.

A provider pick applies to the cloud turns. The small model's own turns travel through a tunnel to our devices, which record only the size and timing of each request. Web searches sit outside the zero-retention requirement either way. How the d1 router picks between our small model and the cloud explains the routing, and What zero data retention covers, and where it stops lists the limits.

If you want zero retention on every cloud message without choosing each time, Self Chat sends each one through OpenRouter with the flag attached and offers only the models that can take it (Self Chat saves your chats only where you choose has more).