Chat with Gemma 4 26B.

The mixture-of-experts Gemma: only 3.8B of its 25B runs per word, so it is much quicker than its size suggests. Reads pictures and video.

Start chatting with Gemma 4 26B now — free, no account

At a glance

  • Maker: Google
  • Released: 2026
  • Parameters: 25B
  • Active per token: 3.8B
  • Context window: 256k
  • Size: Medium
  • Memory: Handles very long documents
  • Good at: Quick answers on a budget
  • Supports: tools, images, zero data retention
  • License: Gemma Terms of Use (Google)
  • Upstream API price: $0.40 per million tokens

Announced in Gemma 4: Byte for byte, the most capable open models.

What is Gemma 4 26B?

Gemma 4 26B is the mixture-of-experts member of Google's Gemma 4 family, and the quickest way to get Gemma quality without paying for it in latency. The trick is in the slug: a4b means roughly four billion parameters active — of the 25B total, only about 3.8B wake up for any given token. That is the same sparsity trick the other fast models here use, applied to a family known for its writing.

In practice the 26B punches well above its active size: it reads pictures and video as well as text, keeps a 256k-token context window — roughly 190,000 words — in view, supports tools, and is covered by zero-data-retention terms. At $0.40 per million tokens, identical to its dense 31B sibling, it is one of the least expensive multimodal models on UsingOpen.

Google's positioning for the pair is unusually honest: same family, same price, different temperaments. The 31B is the careful writer; the 26B is the quick one. If your traffic is conversational — lots of turns, images mixed into the flow, answers that need to be good and prompt rather than perfect — the 26B is the Gemma that fits.

What can you do with Gemma 4 26B?

A cheap, quick multimodal Gemma earns its keep on:

  • Conversational work with multimodal context in the loop when using API endpoints that support image input.
  • Video questions — frames and clips — alongside text in the same chat.
  • High-volume drafting and rewriting where Gemma's prose quality matters but depth does not.
  • Summarising long documents within its 256k-token window, roughly 190,000 words.
  • Light agent workflows with tool support, where cost per turn adds up fast.
  • A budget default for people who want a Google open model without Google flagship prices.

How does Gemma 4 26B compare?

Against the 31B, it is speed against care at the same $0.40: the 31B's dense 31 billion parameters give its answers more polish, while the 26B's 3.8B-active design gets you more answers per minute. Against the bargain bin, it costs a little more than NVIDIA's Lightning ($0.25, text-only, no tools) or DeepSeek's V4 Flash ($0.18, text-only) — and reads pictures and video, which they do not.

Its true peers are the other cheap multimodal models: Ant Group's Ling 3.0 Flash VL matches its frugality with picture and video reading (though with a much smaller context), and Xiaomi's MiMo V2.6 Flash ($0.28) adds sound and a million-token window. If you want Google-flavoured multimodality on a budget, this is the one.

Is Gemma 4 26B right for you?

The 26B is the Gemma for conversational volume: many turns, mixed inputs, questions about pictures and video, and a price — $0.40 per million tokens — that never becomes the story. Its sparse design answers quickly, and Google's writing pedigree shows through even at speed, which is not true of every cheap model. It is also a sensible default for assistants aimed at general audiences: inexpensive enough to run all day, capable enough that users do not notice the corners, multimodal enough to handle whatever people paste in.

Pick the dense 31B instead when polish matters more than pace, or a million-token model when the context, not the clock, is the constraint. If your budget is tighter still, Ling 3.0 Flash VL and DeepSeek V4 Flash cost less, but neither reads pictures and video together at $0.40 — and that combination is the 26B's whole reason for existing. Teams that need ears as well as eyes should look at Xiaomi's MiMo Flash, which adds audio for less.

Frequently asked questions

What does the a4b in Gemma 4 26B mean?

It describes the architecture: about four billion parameters active. The model has roughly 25 billion parameters in total, but only around 3.8B wake up for any given token — a mixture-of-experts design that keeps it much faster than its size suggests. That is the same class of trick as GPT-OSS (5B of 117B) and DeepSeek’s Flash models — the industry has converged on sparsity as the way to make capable cheap.

What is the context window of Gemma 4 26B?

256k tokens — roughly 190,000 words. Long documents, whole books and sizable code modules fit in a single conversation. Both Gemmas share the 256k window, so there is no memory argument within the family — only a speed-versus-polish one.

Does Gemma 4 26B support image input?

Yes — it reads pictures and video as well as text, which is unusual at its price. Its dense sibling, the 31B, reads pictures but not video.

How does Gemma 4 26B compare to Gemma 4 31B?

They cost the same $0.40 per million tokens. The 26B is the quick one — sparse, video-capable, built for volume. The 31B is the careful one — dense, with stronger writing and more deliberation. Same price, different temperaments. The price tie means the decision is temperament, not value: writers gravitate to the dense model, builders of chat-like flows to this one.

Who makes Gemma 4 26B?

Google, as part of its Gemma family of open models — the open-weight line built from the research behind its larger systems, released under Google's Gemma Terms of Use.

Is Gemma 4 26B good value for money?

At $0.40 per million tokens with picture and video input and a 256k window, it is among the cheapest multimodal models on UsingOpen, and the least expensive one from a major Western lab. Western-lab multimodality at $0.40 is unusual; most models with this breadth of input cost two to ten times more.

Is Gemma 4 26B free to chat with on UsingOpen?

Yes. Pick it in the model picker and start chatting — no account, no payment, and your history stays on your device. For anyone building a public-facing assistant on a budget, it is a natural first stop.

Related models on UsingOpen

Also worth a look if you are weighing Gemma 4 26B up:

  • Gemma 4 31B by Google: the dense 31B Gemma — strong writing, picture input and visible thinking.
  • Ling 3.0 Flash VL by inclusionAI: the bargain vision model — 5.5B of 124B awake, the cheapest picture reader on the site.
  • MiMo V2.6 Flash by Xiaomi: the quick Xiaomi model — 15B of 309B awake, reading pictures, video and sound.
  • Nemotron 3.5 Lightning by NVIDIA: the speed-first 30B — just 3B awake per token, built for high-volume text work.

Head-to-head comparisons

See how Gemma 4 26B lines up with the models people most often weigh it against:

Model availability changes as upstream capacity comes and goes — the picker on the chat page shows what is ready right now.

Chat with Gemma 4 26B free on UsingOpen — no account, no sign-up

← All open-weight models · Compare models · Free AI chat · FAQ