Gemma 4 12B
Open weights · Apache 2.0
Released June 2026, an 11.95B dense Gemma 4 model and the first mid-sized Gemma with an encoder-free design that feeds image and audio inputs directly into the transformer. Apache 2.0 license.
No provider we track hosts it. Self-hosting starts at $0.05 an hour: see the estimate.
1 provider
Key Specifications
- Context window
- 256K tokens
- Knowledge cutoff
- Inputs
- Text, Image, Audio Outputs: Text
Gemma 4 12B pricing by provider
| Provider | Input / 1M tokens | Output / 1M tokens | Cost at 10M in + 2M out | |
|---|---|---|---|---|
|
Google Cloud
Creator
|
Open weights, no hosted price | View | ||
Heads up: We do our best to keep these specs & prices accurate. However, cloud costs may fluctuate based on region, usage, and other factors not listed here. These are estimates based on common setups and are for informational purposes only. Always verify current rates & exact specs with the provider before provisioning. LLM rates are base-tier, on-demand prices per 1M tokens; cached-input, batch and long-context tiers are not included.
Compare every model at this volume in the LLM cost calculator.
Capabilities
Function calling
Connect to external tools, APIs, and systems.
Structured output
Return responses in structured formats like JSON.
Estimated cost to self-host Gemma 4 12B
Gemma 4 12B has 11.95B parameters. At 4-bit it needs about 7 GB of GPU memory, at 8-bit 14 GB and at BF16 29 GB, counting 20% on top of the weights for KV cache and runtime overhead.
| Precision | Memory needed | Cheapest rentals that fit | Per month, 24/7 |
|---|---|---|---|
| 4-bit | 7 GB |
GTX 1070
($0.05/hr, tight)
RTX 3060 Ti
($0.07/hr, tight)
|
$36 |
| 8-bit | 14 GB |
RTX 5060 Ti
($0.09/hr, tight)
RTX 4060 Ti
($0.10/hr, tight)
|
$65 |
| BF16 | 29 GB |
RTX 5090
($0.46/hr, tight)
RTX A6000
($0.56/hr)
|
$331 |
Memory is parameters × bytes per weight at each precision, plus 20% for KV cache and runtime overhead. Rentals are the cheapest cards that hold it, at provider-weighted median on-demand prices for the week of August 17, 2026, consumer cards included, in nodes of up to 8 GPUs; a fit with under 15% headroom is marked tight. A month is 720 hours. See the best-value GPUs guide for the same table across models, and cloud GPU pricing for every card.
More from Google Cloud
| Model | Context | Input / 1M | Output / 1M |
|---|---|---|---|
|
Gemini 3.5 Flash-Lite
|
1M | $0.30 | $2.50 |
|
Gemini 3.6 Flash
|
1M | $1.50 | $7.50 |
|
Gemini 3.7 Flash
|
1M | $1.50 | $7.50 |
|
Gemini 2.5 Flash
|
1M | $0.30 | $2.50 |
|
Gemini 2.5 Flash-Lite
|
1M | $0.10 | $0.40 |
|
Gemini 2.5 Pro
|
1M | $1.25 | $10.00 |
|
Gemini 3 Flash Preview
|
1M | $0.50 | $3.00 |
|
Gemini 3 Pro Preview
|
1M | ||
Models near this price
The models nearest this one by input rate, five each way, at each one's cheapest listed provider.
Frequently Asked Questions
How much does Gemma 4 12B cost?
Google Cloud publishes Gemma 4 12B's weights and lists no hosted price. Running it yourself costs whatever the hardware does; see cloud GPU pricing.
Which providers offer Gemma 4 12B?
One provider lists Gemma 4 12B: Google Cloud (open weights, no hosted price).
What is Gemma 4 12B's context window?
Gemma 4 12B accepts up to 256K tokens of input per request. The context window is the prompt plus any documents, conversation history and tool results sent with it; every token in it is billed at the input rate.
What is Gemma 4 12B's knowledge cutoff?
Gemma 4 12B's knowledge cutoff is January 2025: its training data runs up to that month and it has no built-in knowledge of later events.
What inputs and outputs does Gemma 4 12B support?
Gemma 4 12B accepts text, images and audio as input and produces text. Its listed capabilities are function calling and structured output.
Can I self-host Gemma 4 12B?
Yes. Google Cloud publishes Gemma 4 12B's weights under the Apache 2.0 license. At 4-bit it needs about 7 GB of GPU memory; the cheapest rental that holds it is GTX 1070 at $0.05 per hour, about $36 a month. The estimate above prices 8-bit and BF16 too.