Gemma 3 4B
Google · released Feb 20, 2025 · google/gemma-3-4b-it
- Type
- Open weightsGemma Terms of Use
- Params
- 4.3B
- Context
- 131K
about 98K words of context · download allowed, licence restricts use
Our take
Written Aug 3, 2026Gemma 3 is a compact downloadable model from Google that accepts both text and images and returns text. It fits in small hardware footprints and carries a restricted licence that permits commercial use with limits.
Pick this for budget vision-language tasks where hosted inference cost matters, or for lightweight deployment on edge or local hardware with its 4.3 billion parameters. Use it when you need up to 131,072 tokens in a single request at small-model scale. Skip it if you need a fully permissive licence like Apache 2.0, or if you need measured throughput guarantees on every endpoint.
The case for it
- Extremely low hosted inference cost for a vision-language model, with three tracked offers at identical pricing.
- 131,072-token request limit — no smaller-context variant listed in its class.
- Multimodal input at small-model scale: text and images in, text out from 4.3 billion parameters.
The case against it
- Only one of three tracked endpoints lists measured throughput; the rest are unverified in our data.
- Gemma Terms of Use is not fully permissive — redistribution and modification terms are limited versus Apache 2.0 alternatives.
- Arena scores trail top performers by substantial margins, with a 49.6-point gap between its text and maths results showing uneven capability.
How good is it?
IntelligencePuzzles, maths, exam questions
Arena Text (overall)123rd of 143 · 1303.4
CodingWriting and fixing code on its own
Arena Coding140th of 143 · 1273.9
Arena Coding is the only board that has scored it for this.
AgenticPlanning, calling tools, staying on task
Nobody we watch has scored Gemma 3 4B for this. We would take the rating from Arena Agent (IPS).
WritingWe do not rate this
Two boards come close and neither tests writing: Arena Creative Writing asks people which of two replies they prefer, and LiveBench Language tests whether a model understood a passage. So we show where Gemma 3 4B placed and give it no mark out of five.
These tests check whether a model follows instructions — a precondition for all the work above, but not a measure of how well that work is done, which is why they get no rating.
Every published score for this model6 scoresEvery figure we hold, from 6 boards, with who ran it and a link to the source — including the boards no rating above is built on.
Can you run it yourself?
- Fits in memory
- weights load entirely on the card
- Spills to system RAM
- some weights offload; much slower
- Too large
- will not load even with offload
- est
- size is calculated; the verdict could change by 10%
Comfortable fit
GeForce RTX 4090 · 24 GB
Room to spare. 18.7 GB spare means a 10% error in the size would not change the answer.
GeForce RTX 5090 · 32 GB
Room to spare. 26.7 GB spare means a 10% error in the size would not change the answer.
Apple M2 (8-core GPU, 8GB unified) · 8 GB
Room to spare. 1.9 GB spare means a 10% error in the size would not change the answer.
Memory use by level
Against a 24 GB card.
All 3 sizes here are calculated, not measured. We hold no measured file for this model, so each size comes from the parameter count and every verdict above inherits that uncertainty.
Check against your own machine → · All 71 devices, with every size →
Or rent it from someone else
Cheapest of 3 live listings. Picked at the widest standard context we hold, within one quantisation slice, so the numbers beside it are a price one host actually charges.
- per 1M tokens
- $0.050 in / $0.10 out
- Context served
- 131K
- Throughput
- Not measured
| Provider | In / out per 1M tokens | Context | Throughput | Trains on prompts | Logs prompts | Zero retention |
|---|---|---|---|---|---|---|
| OpenRouter | $0.050 / $0.10 | 131K | not measured | Unknown | Unknown | Unknown |
| DeepInfrabf16 | $0.050 / $0.10 | 131K | 24 tok/s | No | No | Confirmed |
| DeepInfrabfloat16 | $0.050 / $0.10 | 131K | not measured | Unknown | Unknown | Unknown |
DeepInfraAcross the 3 listings we hold: 1 say they do not train on prompts, 0 say they do and 2 do not say. 1 appear in the zero-retention registry we check; the rest are unknown to us rather than confirmed either way.
What each host's API supports
From the parameter list each endpoint publishes. Streaming is omitted: nothing we hold reports it, for any model.
- ✓
- Supported
- ✗
- Not supported
- Not published
- host gave no parameter list
| Provider | Tool calling | JSON output | Strict schema |
|---|---|---|---|
| OpenRouter | ✗ | ✓ | ✓ |
| DeepInfrabf16 | ✗ | ✓ | ✓ |
| DeepInfrabfloat16 |
Tool calling: 0 of 3 listings say yes, 2 say no, 1 publishes no parameter list. JSON output: 2 of 3 listings say yes, 1 publishes no parameter list. Strict schema: 2 of 3 listings say yes, 1 publishes no parameter list.
Models people weigh against Gemma 3 4B
When we formed this view
Dates behind this page
Prices last checked 6h ago
What we do not know about this model yet
- We hold no measured file for it, so all 3 sizes on this page are calculated from the parameter count.
- 1 of 3 listings publish no parameter list, so what their API accepts is unknown to us.
- Nothing we hold says whether an endpoint streams, so we do not show it either way.
- 2 of 3 listings do not say whether they train on prompts.
- We hold no cached-input rate for any of its listings.
Licence and identifiers
What the licence allowsGemma Terms of Use, what it allows commercially, and the identifiers you need to pull this model — its Hugging Face repo, our slug and a machine-readable card.
Licence
Gemma Terms of Use
Commercial use allowed, but Google's prohibited-use policy applies and can be updated over time — terms are less static than Apache/MIT.
Identifiers
- Hugging Face
- google/gemma-3-4b-it
- Modality record
- text+image->text
- Catalogue slug
- google-gemma-3-4b