L3
Meta
Llama 3.1
Llama 3.1 is Meta's 8B, 70B and 405B family with 128K context, eight supported languages and downloadable weights.
Open Source
G4
Google DeepMind Advanced
Gemma 4
Gemma 4 is Google's active Apache-2.0 multimodal family with 128K–256K context, 140+ languages, reasoning and function calling.
Large Language ModelText Open Source
What is this model and why does it matter?
Gemma 4 is Google's active Apache-2.0 multimodal family spanning small edge models through 31B and MoE variants with up to 256K context.
Open multimodal assistantsReasoningCodingAgentsEdge AIFine-tuning
Gemma 4: features, use cases and important details
Gemma 4 is Google DeepMind’s current open model family released in March 2026.
Verified model facts
Google documents five main sizes/architectures, multimodal input, configurable thinking, built-in function calling, support for more than 140 languages and context windows from 128K to 256K.
Current status
The models are active under Apache 2.0, and selected 31B/26B A4B instruction models are also hosted through the Gemini API.
Best fit
It is well suited to open multimodal assistants, coding, agents, edge deployment and fine-tuning.
Limitations
Modalities and context differ by size, so users should consult the exact checkpoint instead of treating all Gemma 4 variants as identical.
How to use this model
- Choose E2B, E4B, 12B, 26B A4B or 31B based on hardware and modality needs.
- Download official weights or use supported Gemini API variants.
- Enable thinking when harder reasoning is needed.
- Use function calling for agent workflows.
- Fine-tune for domain tasks when appropriate.
Example prompts
Analyze this image and document together and explain the result.Solve this coding problem using reasoning.Run this agent task and call the required function.
What it can do
- 128K/256K context
- Multimodal input
- Configurable reasoning
- Function calling
- Coding
- 140+ languages
- Open weights
Practical use cases
- Open agents
- Edge AI
- Multimodal assistants
- Coding
- Research
- Fine-tuning
What does it cost?
Open Apache 2.0 weights. Some Gemma 4 variants are hosted through the Gemini API, so usage pricing depends on the exact hosted model rather than one family-wide rate.
Simple summaryWeights are Apache 2.0; local cost depends on model size. Hosted Gemini API pricing applies only to supported Gemma 4 variants.
What stands out
- Apache 2.0 license
- Multiple hardware tiers
- Multimodal
- Long context
- 140+ languages
- Function calling
Things to consider
- Capabilities/modalities vary by size
- No single family-wide hosted price
- Larger variants still need substantial hardware
Important restrictions and trade-offs
- Audio/video support is not present on every size
- Family page must not imply one context value for all models
- Generated output still requires verification
Our editorial take
A major upgrade to the Gemma line: permissively licensed, multimodal and suitable from edge deployment to larger agentic workloads.
