To use the generally available gemini-4-neon model on Google Cloud, confirm enterprise customer eligibility and choose global, us, or eu. For workloads with machine learning processing location requirements, the documentation lists only the us and eu multi-regions.
Suitable tasks: Testing online inference, batch inference, long outputs, or Cloud Storage file inputs in enterprise projects.
Unsuitable tasks: Workloads that depend on the Gemini Live API, RAG Engine, model fine-tuning, or fixed quotas.
Applicable model version: gemini-4-neon, generally available as of 2026-11-18.
Applicable clients, agents, or APIs: Google Cloud Agent Studio and Agent Platform's GenerateContent API, Interactions API, and OpenAI-compatible Chat Completions API.
Recommended thinking levels and settings: Use the default HIGH for an initial test. The model supports LOW, MEDIUM, HIGH, and XHIGH. The official documentation gives no guidance on choosing a level by task.
| Setting | Official value or limit |
|---|---|
| Model ID | gemini-4-neon |
| Model availability | global, us, eu |
| Machine learning processing locations | us, eu |
| Usage options | The documentation lists global, us, and eu for Standard PayGo, Priority PayGo, and Provisioned Throughput. The model also supports batch inference |
| Thinking level | LOW, MEDIUM, HIGH (default), XHIGH. Passing thinking_budget or thinking_level="MINIMAL" returns HTTP 400 |
| Fixed parameters | temperature=1.0, topP=0.95, topK=64, candidateCount=1. The model ignores custom values for temperature, top_p, and top_k |
| Context caching | The model supports implicit caching only, with a minimum of 4,096 tokens. Persistent caching is enabled by default |
To deploy the sample app in Agent Studio, use a Google Cloud project with billing and the Agent Platform API enabled. Deploy the sample app
Confirm that the enterprise customer is verified and that the project you plan to use for the sample deployment has billing and the Agent Platform API enabled.
Choose us or eu if you have a processing location requirement. If you have no such requirement, choose a location listed in the documentation based on your business needs. Select Standard PayGo, Priority PayGo, or Provisioned Throughput based on request and capacity needs.
Send a short test request to gemini-4-neon using the default HIGH. Check the API limits below before passing files or enabling long output.
Use the Interactions API or GenerateContent v1beta1 for long outputs. The Google Gen AI SDK and Agent Studio can continue generation automatically. If you manage the loop yourself, pass continuation_token or use background=true with the Interactions API.
The Google Cloud model page lists these request limits.
| Item | Documented value |
|---|---|
| Shared input and output token limit | 1,048,576. Applies to the GenerateContent, Interactions, and OpenAI-compatible Chat Completions APIs |
| Output limit per foreground segment | 65,536 tokens, including thinking, function calls, and the final answer |
| Long decoding output limit | 1,048,576 − input token count. Applies to the Interactions API and GenerateContent v1beta1 |
| GenerateContent request size | 500 MiB compressed |
| Interactions synchronous or streaming request size | 256 MiB compressed |
| Interactions stored background request size | 9.75 MiB uncompressed when both background and store are true |
| Interactions Cloud Storage URIs | Up to 100 gs:// URIs per request across all modalities, including URIs inherited through previous_interaction_id |
GenerateContent and Interactions accept PDFs, plain text, and the Office and OpenDocument formats listed in the documentation, either inline or through Cloud Storage.
The security controls table lists data residency, CMEK, VPC-SC, and AXT as supported for both online prediction and batch inference. For context caching, it lists support for data residency, VPC-SC, and AXT, and marks CMEK as unsupported.
The generally available release is open only to verified enterprise customers in select regions. Check access within the specific project.
global appears under model availability and usage options. The machine learning processing locations list includes only us and eu.
The GenerateContent gs:// URI limit varies by modality and file count limits. The 100-URI limit applies to Interactions.
The model outputs text only. It accepts images, audio, and video as inputs. The documentation marks the Gemini Live API, RAG Engine, model fine-tuning, latency tiers, and fixed quotas as unsupported.
On 2026-11-19, we opened the Simplified Chinese Google Cloud model page in the international edition of Tabbit and checked the model version, usage options, parameter defaults, and region table. This summary reflects the official configuration. Project access, deployment, and API requests have yet to be tested firsthand.
Gemini 4 Neon