A tour of the 24 models in the picker

The model picker is the button in the composer that names the model you're talking to. Open it on tryingopen.com and, as of October 2026, you get 24 models you can chat with, plus two you can't pick yet.

The note at the top

Before the first group, the picker prints one short note: "23 of these 24 are open: anyone can download and run them. The other is announced as open, with the weights still to come. $ to $$$ is what each model costs to run. Free for you here."

What counts as an open model covers the fine print behind that 23. Qwen3.8 Flash is a special case: it is Alibaba's hosted production model, built on the open-weight Qwen3.8-Flash-Next.

Start here, Quick and light, Big and powerful

The groups have been there since launch day, August 17, 2026, and sort the models by speed and strength. Each has a short hint beside its name.

Start here: "A good first try"

One model lives here: Liquid AI's LFM2.5-VL 3B, the default, labelled "Runs on an edge device". It answers quick questions and pictures on Offline Base's own hardware and hands the rest to a cloud model you choose, GLM 5.3 Flash unless you pick another.

Our default model runs on a palm-sized Jetson Orin Nano covers the hardware, and How the d1 router picks covers the decision.

Quick and light: "Small enough to answer fast"

Seven small and medium models sit here: Gemma 4 E2B, Qwen3.8 27B, Nemotron 3.5 Lightning, Gemma 4 31B, Gemma 4 26B, Muse Glimmer 30B and LFM2.5 24B.

Gemma 4 E2B is the second edge entry, added October 2, 2026. It runs on the same boards, text only, and hands harder questions to Gemma 4 31B unless you pick another cloud fallback. Gemma 4 on tryingopen follows the family from our boards to the cloud. LFM2.5 24B is the newest arrival: it joined the day Liquid AI released it.

Big and powerful: "Slower, and stronger on hard work"

The other 16 usable models are here: GLM 5.3 Flash, Qwen3.8 Flash, DeepSeek V4 Flash, DeepSeek V4 Pro, GPT-OSS 120B, DeepSeek V4.1 Flash, GLM 5.3, Inkling, Hy4 Preview, Mistral Large 4, Nemotron 3 Ultra, Qwen3.8 2.4T, MiMo V2.6 Flash, MiMo V2.6 Pro, Ling 3.0 Flash VL and Inkling Small. MiniMax M3 and GLM 5.2 are listed here too, greyed out.

A model's group follows its total size, which puts some fast models here. GLM 5.3 Flash, the default's cloud fallback, counts as big, though only a small share of it works on each word. What total and active parameters tell you explains the gap.

What Learn more shows

Each row ends with an info button, Learn more. It opens a side panel with:

• a short description, plus the model's size, context window and what it's good at
• chips for what it does here: "Reads images", "Makes files" and "Searches the web"
• what happens to your messages with that model. A zero-retention model's section is headed "Nothing is saved". A model that may log opens with the warning already showing.
• the maker's announcement post when we have one, embedded or as a link that opens in a new tab

The $ / $$ / $$$ bands

Since September 22, 2026, each row shows $, $$ or $$$ beside the name. Hover over it for the label: "Cheap to run", "Mid-priced to run" or "Expensive to run". The band is what the model costs Offline Base to run. You pay nothing, which is what "Free for you here" means.

The $$$ models are Qwen3.8 27B, GLM 5.3, Inkling, Hy4 Preview and Qwen3.8 2.4T. Group follows size and band follows running cost, so Qwen3.8 27B sits in Quick and light with $$$ beside it.

Exact prices change with the provider and the date, while the band stays put. For a reply served through OpenRouter, the info icon under it shows what that reply cost us, as billed. What each reply tells you covers the rest of that panel.

The badges

The edge entries skip the "Provider may log" badge, because their row already says where they run. Their Learn more panel has a section headed "Where your messages go", which walks through each hop, including the cloud turns and web searches that go to outside companies.

Picking by what you need

• You want to send a picture. Thirteen models read pictures, the default among them. Look for the "Reads images" chip. Gemma 4 E2B is text only on our boards. The full list is in Sending pictures and PDFs, and getting files back.
• You want a file made, or a web search. Both run on tools. Every cloud model in the picker has tools except Nemotron 3.5 Lightning, which answers as a plain chat model here, and Gemma 4 26B goes without them when CoreWeave serves it. Search has more gaps. Qwen3.8 Flash never searches, and nothing sent through ngrok's gateway searches either: every "Provider may log" model except Mistral Large 4 when no provider is picked, and Gemma 4 31B, Gemma 4 26B and GPT-OSS 120B on "Let OpenRouter choose", which sends them to Amazon Bedrock. The "Searches the web" chip comes with tool support, so it shows on these models too. How web search works covers searching.
• You have a long document. A chat holds 24,000 characters of text, replies included, so a long document goes in as a PDF. The two edge models have the smallest windows: 8K for Gemma 4 E2B (the most one of our boards holds for a single chat) and 32K for LFM2.5-VL 3B. The cloud models that take 1M include GLM 5.3 Flash, GLM 5.3, Qwen3.8 Flash, DeepSeek V4 Flash, DeepSeek V4 Pro, DeepSeek V4.1 Flash, Hy4 Preview, Qwen3.8 2.4T, MiMo V2.6 Flash and MiMo V2.6 Pro. Each Learn more panel lists the figure.
• You want zero retention. A cloud model without the "Provider may log" badge goes only to providers that promise zero data retention. For five of the nine badged models, picking a provider in the composer does the same; the other four have no zero-retention provider yet. Web searches and PDFs sit outside the promise. What zero data retention covers has the details.

The list changes often, and What's new in the picker since September tracks what came and went. Once you find a setup you like, a link can carry it.