Gemma 4 E2B — Multimodal Language Models AI Model
The compact Gemma 4, E2B, runs in bfloat16 on about 10GB of VRAM and still handles visual Q&A, captioning and image comparison with a 128K context. Pick it when the 31B model is more than the job needs.
Also runs in From Receipt to Picture · One Answer, Checked by Three AIs
What it does
Gemma 4 E2B is the smallest model in Google DeepMind's Gemma 4 family. The E stands for effective: it is built to perform above what its size suggests, and it answers questions about text and images across a 128K-token context.
It runs in bfloat16 on roughly 10GB of VRAM, which is what makes it the one to pick when the 31B Gemma 4 is more than the job needs. Two images can go in together, so a comparison does not have to be stitched into one picture first.
When to use it
- Asking a question about an image and getting a text answer
- Captioning, or describing what is going on in a picture
- Comparing two images in one pass, since both go in together
- Long documents and transcripts, on a 128K-token context
- Turning reasoning on for a hard question and off for a quick one
Details
- CategoryMultimodal Language Models
- Year2026
- LicenseApache-2.0
Compliance & Provenance
- ProviderGoogle (open weights) · GPAI
- EU AI Act RiskLimited Risk
- Art. 50 TransparencyRequired — AI-generated outputs are marked
Inputs & Outputs
- TextInputstring
Text prompt or question
- Image 1Inputimageoptional
Optional reference image for visual understanding. When provided, the model analyzes the image content together with the text prompt. Supports up to 4096x4096 resolution. Connect from any image-producing node (e.g. super-resolution, deblur, inpainting) to analyze processed results.
- Image 2Inputimageoptional
Optional second reference image. When provided alongside Image 1, the model can compare, contrast, or jointly reason over both images with the text prompt. Supports up to 4096x4096 resolution.
- TextOutputstring
Generated text response
Parameters
- max_new_tokensintegerdefault 2048range 1–8192advanced
Max new tokens
Maximum tokens the model may generate in its response. Default 2048. Lower to constrain length / cost; raise (up to 8192) for long-form output. Generation also stops at natural EOS.
- temperaturefloatdefault 1range 0–2advanced
Temperature (0-2)
Sampling temperature. 0.0 = deterministic. 1.0 (Gemma default) = balanced. >1.0 = more divergent. Lower for factual Q&A, higher for creative writing.
- top_pfloatdefault 0.95range 0–1advanced
Top-p (nucleus)
Nucleus sampling threshold — keeps the smallest token set whose cumulative probability ≥ top_p. 0.95 (default) is a strong default. Lower values produce more focused output.
- top_kintegerdefault 64range 1–100advanced
Top-k
Top-k sampling — restricts the candidate set to the k highest-probability tokens. 64 (Gemma default) is permissive. Lower for more deterministic output.
- enable_thinkingbooleandefault Offoptions: On, Offadvanced
Thinking mode (On/Off)
When 'On', Gemma 4 runs an internal chain-of-thought before answering — better quality on multi-step reasoning and complex visual analysis, at the cost of higher latency and token count. Default 'Off' for fast simple replies. Switch on for benchmarks or hard problems.
Tags
Alternatives in Multimodal Language Models
- Agnes-3.0-Flash
Agnes 3.0 Flash Preview — 33B hybrid-attention vision-language model, 262K context, bf16. Open-weights preview checkpoint, distinct from the production Agnes 3.0 Flash.
- Gemma 4 31B
Gemma 4 31B Dense flagship VLM, 256K context, thinking mode.
- Mage-VL
Mage-VL — 4B codec-native streaming multimodal model from Microsoft. A from-scratch Mage-ViT visual encoder on a Qwen3-4B decoder, trained to read images and video through codec structure rather than uniform frame sampling. This entry serves its image-and-text mode; its video and proactive-streaming modes are not wired up here.
- Qwen3.6-35B-MoE
Qwen 3.6 35B/A3B MoE VLM, 256K context. fast inference.
- Qwen3.8-27B
Qwen 3.8 27B Dense VLM, 256K context, bf16.