Throughput (tokens / sec)
avg. 215tokens/s
Gemma 4 quality at 4B-active speed
Run instantly. Pay only for what you use.
Gemma 4 26B A4B is the sparse MoE member of Google DeepMind's Gemma 4 family, released under Apache 2.0: 25.2B total parameters with just 3.8B active per token across 128 fine-grained experts, a 256K-token context window, and native text + image input, pre-trained across 140+ languages.
It keeps near-flagship quality at small-model compute: within about 3 points of the dense Gemma 4 31B on MMLU-Pro and MMMU-Pro, with a configurable thinking mode that adds step-by-step reasoning only where a request needs it.
With thinking enabled it posts 82.6% on MMLU-Pro, 88.3 on AIME 2026, and 73.8 on MMMU-Pro — the speed-and-cost pick of the Gemma 4 line, with strict function calling and JSON output for production pipelines.
3.8Bactive of 25.2B total
128 fine-grained experts route per token — near-31B quality at small-model compute.
256Ktoken context window
Holds long documents, retrieved passages, and multi-turn sessions in one prompt.
140+languages in pre-training
35+ supported out of the box — one deployment serves a global user base.
Gemma 4 26B A4B is strongest where volume meets variety — multilingual assistants, visual document pipelines, and high-throughput extraction that need flagship-adjacent quality without flagship cost.
82.6% on MMLU-Pro and an Arena ELO within a dozen points of Gemma 4 31B — from a fraction of the compute per token.
73.8 on MMMU-Pro — screenshots, PDFs, and diagrams are read alongside text, no separate vision model to route to.
Step-by-step reasoning on demand — 88.3 on AIME 2026 with thinking enabled, fast direct replies with it off.
Strict function calling and JSON mode — tool arguments and extraction pipelines that parse the first time.
Multilingual chat and support at scale
35+ languages out of the box, 140+ in pre-training — one deployment serves a global user base at 4B-active cost.
Multimodal document and screenshot pipelines
Native vision at 73.8 on MMMU-Pro reads screenshots, PDFs, and diagrams alongside text in one model.
High-volume fleets on a small-model budget
Arena-ranked within a dozen ELO points of the 31B flagship while activating 3.8B parameters per token.
Document and image extraction to JSON
Function calling and JSON mode return schema-valid output for ticket, form, and invoice pipelines.
The hardest coding and reasoning ceiling
The dense Gemma 4 31B scores higher across the board — step up to it when the last few points matter.
Tight output-token budgets with thinking on
Thinking mode is verbose — about 2× the median output tokens in Artificial Analysis testing. Cap or disable thinking for cost-sensitive replies.
Audio input or speech understanding
Inputs are text and image only. Transcribe audio upstream before sending it to the model.
avg. 215tokens/s
avg. 540ms
82.6%MMLU-Pro
Graduate-level knowledge and reasoning across 14 domains.
73.8MMMU-Pro
University-level reasoning over images, charts, and diagrams.
88.3AIME 2026
Competition mathematics measuring multi-step reasoning.
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.
Multimodal long-context work with thinking control
Tool-heavy coding at 3B-active cost
Low-latency reasoning on a tiny budget