Apache 2.0 from Google, weights free, nothing to accept. This is the rare model in the catalog you can genuinely run on hardware you already own: 2.3 billion effective parameters, built for phones and laptops, so a self-host needs the Hugging Face transformers library and not much else. In your browser it runs on CNAPS Studio, included in the basic plan.
The reason to care is what it takes in. Most models this size read text and maybe images. This one also listens, handling up to thirty seconds of speech for transcription or translation without a separate speech pipeline in front of it, and it reads up to sixty seconds of video as one frame per second. Text, pictures and audio can be mixed in any order inside a single prompt.
The cost is ceiling. On advanced maths it scores 37.5 out of 100 where its larger sibling scores 89.2, so this is a model for everyday work rather than hard reasoning. Three multimodal models sit in the CNAPS Studio catalog, and this is the only one that hears.
What it actually is
A dense transformer that alternates local and global attention, which is the trick that keeps memory manageable at a 128,000 token context. The E stands for effective: 2.3 billion parameters do the work, and 5.1 billion is the figure including the embedding tables, which is why you will see both numbers quoted.
Four input types are supported, and audio is the distinctive one. Speech recognition and speech translation both work up to thirty seconds per clip. Video is read as sequential frames at one per second up to a minute. Function calling is native, so it can be pointed at external tools, and there is a thinking mode that makes it reason step by step before answering.
Five settings matter in CNAPS Studio. Max new tokens caps the answer length, and 1,024 suits most tasks. Temperature controls invention: 0 for deterministic factual work, 1.0 is Google's general recommendation. Top-P at 0.95 and Top-K at 64 keep word choice sensible. And the thinking toggle is the real lever: on for anything hard, off for conversation, because it costs time.
Effective parameters2.3 billion, or 5.1 billion including embeddingsContext window128,000 tokens, roughly 100,000 wordsModalitiesText, image, audio, videoAudioSpeech recognition and translation, up to 30 secondsVideoUp to 60 seconds at one frame per secondThinking modeBuilt in, on or offFunction callingNativeLanguages35 or more out of the box, pre-trained on 140 or moreLicenseApache 2.0Downloads last month3,235,641
The published numbers
Google publishes a full scorecard, and every figure below is a percentage out of 100 unless noted. On MMLU Pro, which tests broad expert knowledge, it scores 60.0. On GPQA Diamond, graduate-level science questions, 43.4. On LiveCodeBench v6, real coding problems, 44.0. On AIME 2026, competition mathematics, 37.5. On MMMLU, the same knowledge test across languages, 67.4. On MMMU Pro, which needs reading images and text together, 44.2.
Two audio numbers use different scales and are worth reading carefully. FLEURS scores 0.09 for speech recognition, and there lower is better, since it counts mistakes rather than correct answers. CoVoST scores 33.47 for speech translation, on a scale where higher is better.
All of these come from Google, so they are vendor-reported. The counterweight is unusually strong here: these are public benchmarks that independent groups re-run constantly, and adoption is enormous at 3,235,641 downloads in the last month. A vendor claim on a public leaderboard that millions of people can check is a different kind of claim from a private evaluation.
How it compares to the other multimodal models here
Gemma 4 E2BGemma 4 31BQwen3.6-35B-A3BSize2.3B effective, runs on a phone or laptop30.7B, needs a data-centre GPUSee its own postContext128,000 tokens256,000 tokensSee its own postAudio inputYes, up to 30 secondsNoSee its own postMMLU Pro60.085.2See its own postAIME 202637.589.2See its own postGPQA Diamond43.484.3See its own postLicenseApache 2.0Apache 2.0See its own post
The two Gemma columns are directly comparable, which is rare in this series: same publisher, same benchmarks, same reporting. The gaps are large and in the expected direction. What the table does not show is the thing that matters most in practice, which is that one of these runs on a phone and the other does not.
Reach for Gemma 4 E2B when the input includes audio, when the work is everyday summarising, describing or classifying, or when running locally and privately matters. Reach for Gemma 4 31B when the task is hard reasoning, competition-grade maths or code, or documents too long for a 128,000 token window. Reach for Qwen3.6-35B-A3B when you want a second opinion from a different publisher.
What to chain it with
It takes text, images, audio and short video, and returns text. Being the only model here that accepts audio makes it the natural endpoint for anything spoken.
Two chains earn their place. For meeting or interview notes, Speech Recognition (Clips) handles the long transcript and this model summarises it, or for anything under thirty seconds it does both jobs itself in one step. For images, any detector or captioner in the catalog feeds it: DeepSeekOCR reads a page and this model answers questions about what was on it, or BLIP (Image Description) describes a photo and this model turns a batch of descriptions into a report.
Open Gemma 4 E2B in CNAPS Studio, hand it a thirty-second voice memo with no transcript step in front, and see whether it needs one.
Sources
- huggingface.co/google/gemma-4-E2B-it for the Apache 2.0 license, the parameter counts, the 128,000 token context, the supported modalities, the audio and video limits, every benchmark score quoted above, and the monthly download count.
- huggingface.co/google/gemma-4-31B-it for the comparison column.
- docs.cnaps.ai/ai-model-hub for the CNAPS Studio settings, their defaults and recommended values.