tryingopen: try any open model, no account

tryingopen is Offline Base's free site for trying open models. Its homepage says "Try any open model." and promises "No account, nothing saved: pick a model, start chatting, and share any chat as a link."

The models

On October 14, 2026, the picker listed 24 chat models. Two run on our own NVIDIA Jetson boards: Liquid AI's LFM2.5-VL 3B, the default pick, and Google's Gemma 4 E2B. The other 22 run at cloud providers:

Maker	Models
Z.ai	GLM 5.3, GLM 5.3 Flash
Alibaba	Qwen3.8 Flash, Qwen3.8 27B, Qwen3.8 2.4T
DeepSeek	DeepSeek V4 Flash, V4 Pro, V4.1 Flash
Google	Gemma 4 31B, Gemma 4 26B
NVIDIA	Nemotron 3.5 Lightning, Nemotron 3 Ultra
Thinking Machines	Inkling, Inkling Small
Xiaomi	MiMo V2.6 Flash, MiMo V2.6 Pro
Others	OpenAI's GPT-OSS 120B, Meta's Muse Glimmer 30B, Tencent's Hy4 Preview, Mistral Large 4, inclusionAI's Ling 3.0 Flash VL, Liquid AI's LFM2.5 24B

The picker labels 23 of the 24 as open. The 24th, Mistral Large 4, is announced as open with weights still to come. Of the 23, all but Qwen3.8 Flash have public weights that anyone can download and run. Qwen3.8 Flash is a hosted production model with no published weights, and we include it on purpose.

Each model shows a cost band from $ to $$$ (what it costs us to run) beside the note "Free for you here." While we add capacity, a chat holds up to 12 of your messages.

The second edge entry runs Gemma 4 E2B, text only, and hands harder messages to Gemma 4 31B by default. If you pick a cloud model outright, your first two questions also go to a Jetson, and the Jetson's answers wait behind a tab for comparison.

What the site keeps

Per the privacy policy, we never log the text of messages or replies. We keep an hourly message count per IP address, which expires within about two hours, and cookie-free page-view counts. The chat sets no cookies.

Requests through OpenRouter require zero data retention: if no provider promises it, the request fails. Nine of the 22 cloud models lack a zero-retention route by default and are labeled "Provider may log". For five of them, including GLM 5.3 Flash, choosing a specific provider in the composer restores the requirement. Web searches and PDF parsing sit outside it. The default edge pick is also outside it: its cloud turns go to GLM 5.3 Flash, which may log unless you pick a provider. Our own boards record only the size and timing of requests.

Share any chat as a link

The share button compresses the conversation into the link itself, after the "#", a part of the address that browsers never send to a server. We hold no copy; anyone with the link can read the chat, minus attachments and reasoning.

Since October 13 a copied address also shares your model, provider and cloud fallback.

The default: a 3B model on a Jetson, with a router in front

The default pick is tagged "Runs on an edge device". It answers with LFM2.5-VL 3B, Liquid AI's vision-language model, quantized to Q4_K_M (a mostly 4-bit mix) and running on Jetson Orin Nano boards in our fleet. Each palm-sized board has 8 GB of memory shared by its CPU and GPU.

Each message is read first by d1-3B, a decision model Liquid AI released as open weights on October 7. It answers yes/no, multiple-choice and rating questions about a message, with probabilities, in a single forward pass. We run it with llama.cpp on one of our Jetsons, so the message stays on our hardware while d1 decides where it goes.

When we checked on October 14, d1 answered five yes/no questions per message: code; multi-step math; medical, legal or financial expertise; an unhappy user; and live web information. The small model's chance of answering well is one minus the largest of the first four. At 50% or more a Jetson answers; below that, your chosen cloud model does, GLM 5.3 Flash by default.

When the web signal reaches 50%, the page first runs one web search (Exa, through OpenRouter) and gives the results to the small model to cite, so a weather lookup can stay on the board.

A few rules keep the page answering:

• A PDF goes straight to the cloud.
• If no board with the model is free, the cloud answers and d1 is skipped.
• If d1 fails or takes over 3 seconds, the turn stays on the board.
• If the fleet errors or takes over 4 seconds, the cloud answers, so the page works with every Jetson unplugged.

In one live check on October 14, "Hi! How are you today?" stayed on a Jetson at 98.6%, decided in 626 ms measured from the site. Its web signal came back at 51.6%, just over the line, so the page searched the web for a greeting.

A stricter rule is in the code and rolling out. It asks four questions, starting with whether the message is plainly simple, and keeps a turn on the Jetson only at 70% or more. Borderline messages go to the cloud. On our 37 labeled test chats it scored 37 of 37. The six-question version it replaces in the code (the five above plus a request for a bigger model) scored 30 of 37.

Every edge reply draws its route through d1, with a bar for each signal.

Self Chat, the developer API and closed-model comparison

• Self Chat saves each chat as files in your Google Drive or a folder on your computer, uses zero-retention providers only, and can talk from your browser to a local Ollama, LM Studio, llama.cpp or Jan.
• The developer API is OpenAI-compatible, with 8 text models, 25 credits a day (one credit is $0.01 of model cost) and a key that is a name you pick. It stores your account and keeps usage records for up to 90 days. We never store your messages, though a provider marked "Provider may log" may.
• Compare with a closed model sends a chat, only when you press it, to Claude Sonnet 5 or GPT-5.6 Terra under their makers' own terms.

Offline Base and Gemma Gemini

Offline Base, the early-stage company behind Gemma Gemini, runs tryingopen (more on the company). It began as the open models page on offlinebase.com, which still advertises 10 open-weight models. Its default pick reflects our bet that small models can handle most everyday questions.

Gemma Gemini's first router, built on October 3, was modeled on tryingopen's. Both apps now route on our own hardware with different models, and both can answer with the same Gemma 4 E2B build. From tryingopen to Gemma Gemini tells that story in full.