Gemma 4 on tryingopen, from our boards to the cloud

You can pick Google's Gemma 4 in three sizes on tryingopen. The smallest runs on palm-sized computers we own. The two larger ones answer from the cloud. As of October 2026, all three sit in the picker's "Quick and light" group.

What Google released

Google announced Gemma 4 on April 2, 2026, under what it calls a "commercially permissive Apache 2.0 license". The announcement's headline is "Byte for byte, the most capable open models". It launched four sizes:

• E2B and E4B, edge models with an effective 2B and 4B footprint
• 26B, a mixture-of-experts (MoE) model with 3.8B parameters active per token
• 31B, a dense model

The two edge models have a 128K-token context window, and the 26B and 31B go up to 256K. Google says all four process images and video, and that the edge models also take audio.

Google says the 31B ranked #3 and the 26B ranked #6 among open models on the Arena AI text leaderboard as of April 1, 2026.

On June 3, 2026, Google added a fifth size, Gemma 4 12B, also under Apache 2.0. Google places it between E4B and the 26B and calls it its "first mid-sized model to feature native audio inputs." Its model card lists a dense model with 11.95B parameters and a 256K context.

We offer three of the five: E2B, 26B and 31B. E4B and 12B are not in the picker. For what Apache 2.0 and the other licenses on the site allow, see What counts as an open model, and what the licenses allow.

The three sizes on tryingopen

Gemma 4 E2B on our boards

Gemma 4 E2B joined the picker on October 2, 2026, as our second edge-device entry. The first, and still the default pick, is Liquid AI's LFM2.5-VL 3B. The picker lists the maker of both entries as Offline Base.

The E2B model card gives 2.3B effective parameters, or 5.1B counting its per-layer embeddings. The card says per-layer embeddings give "each decoder layer its own small embedding for every token." What total and active parameters tell you about a model explains how one model ends up with two sizes on its card.

On the boards, E2B is stored at 4-bit precision, about 3.2 GB, and served by llama.cpp on the board itself. Each board is an NVIDIA Jetson Orin Nano with 8 GB of memory. Our default model runs on a palm-sized Jetson Orin Nano covers the hardware in more detail.

Running it on a board changes two things from the card:

• Text only. The card lists text, image and audio input. Our boards run it without a vision projector, so pictures and audio are off here. To send a picture, pick LFM2.5-VL 3B or one of the cloud Gemmas.
• 8K context. The card gives 128K tokens. Here a chat gets 8K tokens, the most one slot on a board holds. The site labels it "Best for short chats".

How d1 routes it

d1, Liquid AI's small decision model, routes E2B the same way it routes the default pick. It reads each message first on one of our boards: quick questions stay on E2B, and anything more, such as code, maths, advice or a longer, detailed answer, goes to a cloud model. Every reply says which model answered.

How the d1 router picks between our small model and the cloud has the full rule.

The cloud fallback is Gemma 4 31B

When d1 sends an E2B turn to the cloud, it goes to Gemma 4 31B by default, so a hard question moves from the smallest Gemma 4 we run to the largest. Since October 8, 2026, you can choose another model in the "Cloud fallback" picker, and a provider picker beside it applies to whichever fallback you choose. Choosing the company that runs your model explains the provider picker.

The edge entries sit outside the site's zero-retention promise. Messages reach our boards through a tunnel, and cloud turns go to outside providers. The boards log three things for each request: the model, the token counts and the timing. Message and reply text stays out of the log. The four things tryingopen keeps about your visit has the full list.

Gemma 4 31B and 26B in the cloud

The two larger sizes answer from the cloud, and the picker lists a 256K context for both. A chat opened directly on either one, from a link for example, starts on CoreWeave through OpenRouter under a zero-retention requirement, and the provider button lists the other companies you can pick. The button's first row, "Let OpenRouter choose", works differently for these two: it sends them through ngrok's AI gateway to Amazon Bedrock. The home page starts on that row, so switching to either model from there keeps it. Both read pictures you attach, so they are the Gemmas to pick for an image question.

• Gemma 4 31B is the dense model, with 30.7B parameters per its model card.
• Gemma 4 26B is the MoE model. Its model card gives 25.2B parameters in total, with 3.8B active per token: each token goes through 8 of its 128 experts plus 1 shared expert. The picker rounds it to 25B, and the site says it is "much quicker than its size suggests."

Both are zero-retention picks. What zero data retention covers, and where it stops has the details. Self Chat sends its cloud messages straight to OpenRouter under zero retention and never uses ngrok's gateway. It leaves both Gemmas out of its picker.

Both models are among the eight models listed for the tryingopen API. The listed prices are 75 credits per million input tokens and 115 per million output tokens for the 31B, and 15 and 60 for the 26B. The API is text only. As of October 14, 2026, the model list and docs are public, and the Keys tab says accounts aren't available yet.

Gemma 4 E2B on Gemma Gemini

Gemma Gemini is a sister chat from Offline Base, first built on October 3, 2026. It sends each message either to Gemma 4 E2B on our Jetson Orin Nano boards or to Gemini 4 Argon through OpenRouter. There, E2B runs as a 4-bit Q4_0 build under Ollama and answers at about 25 to 29 tokens per second.

Its serving software and router differ from tryingopen's, so those speeds describe Gemma Gemini only.

Picking a Gemma

To pair Gemma 4 E2B with the 26B, pick E2B, then choose Gemma 4 26B in the "Cloud fallback" picker.

Since October 13, 2026, the page address carries your setup, including the cloud fallback, so you can share Gemma 4 E2B with the 26B behind it as a link or save it as a preset in your browser. Sharing a chat, a model setup or a preset as a link shows how.