Gemma 4 quality at 4B-active speed
Multimodal long-context work with thinking control
Selected Gemma models served through an optimized inference stack built for lower request cost at production volume.
Infrastructure designed for sustained throughput, stable latency, efficient GPU utilization, and cost-efficient inference.
Prompts and outputs are processed for inference, not stored after completion, and never used to train models.
Use $25 free credits to evaluate throughput, TTFT, latency, request success, and output quality before paying anything or moving traffic.
What is Gemma good for? Gemma is Google DeepMind's family of Apache-2.0 open-weight models — natively multimodal, pre-trained across 140+ languages, and topping open-model user-preference rankings — with a sparse-MoE variant that activates just 4B parameters per request for near-flagship quality at small-model cost.
Out-of-the-box support for 35+ languages and pre-training across 140+ — one deployment serves a global user base, with thinking control for escalated cases.
Native vision reads screenshots, PDFs, and diagrams alongside text — grounded answers over 256K tokens of context.
Gemma 4 31B posts 80% on LiveCodeBench and 89% on AIME — dense-model consistency for code review and math-heavy work.
The 26B A4B MoE activates 4B parameters per token and ranks within a few Arena ELO points of the 31B flagship — near-flagship quality at a fraction of the cost.
Function calling and structured output return schema-valid JSON for ticket, PDF, and form pipelines at production volume.
A 256K context fits retrieved passages and full documents, and configurable thinking adds reasoning depth only where a query needs it.
Code examples
curl -X POST "https://api.entrim.ai/v1/chat/completions" \ -H "Content-Type: application/json" \ -H "Authorization: Bearer $ENTRIM_API_KEY" \ -d '{ "model": "google/gemma-4-26B-A4B-it", "messages": [ { "role": "user", "content": "Why do they call it a building if it is already built?" } ] }'SDKs & docs
Use familiar SDK patterns with quickstart examples for common production setups.
OpenAI-compatible
Keep the request format your team already knows.
No cold starts
Production requests are served without model spin-up delays.
Prompts and outputs are processed for inference, not stored after completion, and never used to train models.
Inference and request processing stay within Entrim’s EU-operated environment, with DPA support for customers.
Requests are encrypted, access is isolated by API key, and traffic is handled with tenant-level separation.
Controls are being built around recognized security frameworks, with limited operational metadata used for billing, monitoring, latency, and abuse prevention.