Gemma 4 31B
Google · released Mar 11, 2026 · google/gemma-4-31B-it
- Type
- Open weightsApache License 2.0
- Params
- 32.7B
- Context
- 262K
about 197K words of context
Our take
Written Aug 3, 2026Gemma 4 is a 32.7-billion-parameter text, image and video model from Google with a permissive Apache licence. It scores consistently across six Arena categories, with coding as its relative standout, and is available from ten hosts with a wide spread in price and speed.
Pick this for general multimodal inference where Apache licensing matters, or for coding workloads where its Arena score is strongest. It suits budget-conscious hosting with faster throughput, or long-context tasks up to 262,144 tokens. Skip it if you need web-development coding specifically, where it lags its own general coding score by a wide margin, or if you want a measured efficiency story — active parameter count is undisclosed.
The case for it
- Broad benchmark coverage with consistent mid-table scores across six Arena categories, all between 1363 and 1498 Elo.
- Coding is the relative standout, 47.2 points above its own overall text score.
- Apache 2.0 licence allows commercial use, fine-tuning and redistribution.
- Ten current offers with meaningful price and speed spread: input rates vary by a factor of six and throughput by nearly four times.
The case against it
- Web development coding lags general coding by 134.7 points, its widest category gap.
- Creative writing is the weakest measured category, 77.6 points below its coding score.
- No efficiency story: with 32.7 billion total parameters and no disclosed active count, every token processes the full model.
How good is it?
IntelligencePuzzles, maths, exam questions
Arena Text (overall)33rd of 143 · 1450.9
CodingWriting and fixing code on its own
Arena Coding36th of 143 · 1498.1
AgenticPlanning, calling tools, staying on task
Arena Agent (IPS)36th of 36 · −0.161
Arena Agent (IPS) is the only board that has scored it for this.
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 4 31B 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 model8 scoresEvery figure we hold, from 8 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%
GeForce RTX 4090 · 24 GB
Loads, but do not expect an assistant. Some of the weights sit in ordinary system memory, which is far slower than the card.
Comfortable fit
GeForce RTX 5090 · 32 GB
Room to spare. 6.4 GB spare means a 10% error in the size would not change the answer.
Apple M3 Pro (18-core GPU) · 36 GB
Room to spare. 2.6 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 22 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
- $0.10 in / $0.34 out
- Context served
- 262K
- Throughput
- Not measured
| Provider | In / out per 1M tokens | Context | Throughput | Trains on prompts | Logs prompts | Zero retention |
|---|---|---|---|---|---|---|
| CoreWeavebf16 | $0.10 / $0.34 | 262K | 39 tok/s | No | No | Confirmed |
| OpenRouter | $0.10 / $0.34 | 262K | not measured | Unknown | Unknown | Unknown |
| DeepInfraturbo tierfp4 | $0.090 / $0.34 | 262K | 50 tok/s | No | No | Unknown |
| DeepInfrafp4 | $0.090 / $0.34 | 262K | not measured | No | No | Unknown |
| OpenInferencebf16 | $0.10 / $0.35 | 262K | 54 tok/s | No | No | Confirmed |
| Venice AIbf16 | $0.12 / $0.36 | 256K | 38 tok/s | No | No | Confirmed |
| Chutesfp4 | $0.12 / $0.37 | 131K | 14 tok/s | No | Yesunknown period | Unknown |
| DeepInfrafp8 | $0.13 / $0.38 | 262K | not measured | Unknown | Unknown | Unknown |
| DeepInfrafp8 | $0.13 / $0.38 | 262K | 49 tok/s | No | No | Confirmed |
| Novita AI | $0.14 / $0.40 | 262K | not measured | Unknown | Unknown | Unknown |
| Novita AIbf16 | $0.14 / $0.40 | 262K | 16 tok/s | No | No | Confirmed |
| Morphfp4 | $0.14 / $0.40 | 175K | 16 tok/s | No | No | Confirmed |
| SiliconFlowfp8 | $0.13 / $0.40 | 262K | 26 tok/s | No | No | Confirmed |
| Crusoe | $0.14 / $0.40 | 262K | 34 tok/s | No | No | Confirmed |
| Parasailfp8 | $0.15 / $0.40 | 262K | 30 tok/s | No | No | Confirmed |
| Friendli | $0.14 / $0.40 | 262K | 111 tok/s | No | Yesunknown period | Unknown |
| Phala | $0.15 / $0.46 | 262K | 21 tok/s | No | No | Confirmed |
| ModelRunfp4 | $0.22 / $0.55 | 262K | 61 tok/s | No | No | Confirmed |
| Together AI | $0.28 / $0.86 | 262K | 18 tok/s | No | No | Confirmed |
| SambaNova | $0.38 / $1.15 | 131K | 121 tok/s | No | No | Confirmed |
| SambaNova | $0.38 / $1.15 | 131K | not measured | Unknown | Unknown | Unknown |
| Cerebrasfp16 | $0.99 / $1.49 | 131K | 29 tok/s | No | No | Confirmed |
DeepInfra
Novita AI
Morph
Together AI
SambaNova
CerebrasAcross the 22 listings we hold: 18 say they do not train on prompts, 0 say they do and 4 do not say. 14 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 |
|---|---|---|---|
| CoreWeavebf16 | ✓ | ✓ | ✓ |
| OpenRouter | ✓ | ✓ | ✓ |
| DeepInfraturbo · fp4 | ✗ | ✓ | ✓ |
| DeepInfrafp4 | ✗ | ✓ | ✓ |
| OpenInferencebf16 | ✓ | ✓ | ✓ |
| Venice AIbf16 | ✓ | ✓ | ✓ |
| Chutesfp4 | ✓ | ✓ | ✓ |
| DeepInfrafp8 | |||
| DeepInfrafp8 | ✓ | ✓ | ✓ |
| Novita AI | |||
| Novita AIbf16 | ✓ | ✓ | ✓ |
| Morphfp4 | ✓ | ✓ | ✓ |
| SiliconFlowfp8 | ✓ | ✓ | ✓ |
| Crusoe | ✓ | ✓ | ✓ |
| Parasailfp8 | ✓ | ✓ | ✓ |
| Friendli | ✗ | ✓ | ✓ |
| Phala | ✓ | ✓ | ✓ |
| ModelRunfp4 | ✓ | ✓ | ✓ |
| Together AI | ✗ | ✓ | ✗ |
| SambaNova | ✓ | ✗ | ✗ |
| SambaNova | |||
| Cerebrasfp16 | ✓ | ✓ | ✓ |
Tool calling: 15 of 22 listings say yes, 4 say no, 3 publish no parameter list. JSON output: 18 of 22 listings say yes, 1 says no, 3 publish no parameter list. Strict schema: 17 of 22 listings say yes, 2 say no, 3 publish no parameter list.
Models people weigh against Gemma 4 31B
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.
- 3 of 22 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.
- 4 of 22 listings do not say whether they train on prompts.
Licence and identifiers
What the licence allowsApache License 2.0, 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
Apache License 2.0
Fully permissive: commercial use, redistribution, and derivatives allowed. Requires attribution and a copy of the license. Includes an express patent grant.
Identifiers
- Hugging Face
- google/gemma-4-31B-it
- Architecture
- Dense
- Modality record
- text+image+video->text
- Catalogue slug
- google-gemma-4-31b