MOUNTAIN VIEW, Calif., April 2, 2026 — Google has introduced Gemma 4, a new family of open models with open weights that is clearly aimed at bringing more capable AI onto local hardware. Released under the Apache 2.0 license, the Gemma 4 family includes four sizes: E2B, E4B, 26B A4B MoE and 31B Dense. Google is positioning the family around high intelligence-per-parameter, multimodal understanding and agentic workflows, with deployment targets ranging from mobile devices and laptops to workstations and edge systems.
For edge developers, the center of gravity is the smaller E2B and E4B models. Google says these variants are optimized for efficient local execution on laptops and mobile devices, and they use per-layer embeddings to improve parameter efficiency for on-device deployments. Both support 128K-token context windows, while the 26B A4B MoE and 31B Dense models extend to 256K. All Gemma 4 models handle text and image input and can analyze video as sequences of frames, while the E2B and E4B models also support audio input. That combination makes the small models relevant for edge applications that need local multimodal perception plus longer-context reasoning.
For computer vision engineers, the most notable part of the release may be the explicit support for image-centric tasks. In its model card, Google lists object detection, document and PDF parsing, screen and UI understanding, chart comprehension, OCR, multilingual OCR, handwriting recognition and pointing among Gemma 4’s image-understanding capabilities. The models also support variable image aspect ratios and configurable visual token budgets, giving developers a way to trade off detail against compute and latency. That makes Gemma 4 more than a text model with vision bolted on; it looks like a plausible building block for document-understanding systems, multimodal field tools and other edge products that need local perception plus reasoning.
Google is also pairing Gemma 4 with a deployment stack aimed at real edge use. In its Google AI Edge announcement, the company says LiteRT-LM adds constrained decoding for structured outputs, dynamic context handling across CPUs and GPUs, and a minimal-memory path for smaller Gemma 4 models. Google says E2B can run in under 1.5 GB of memory on some devices using 2-bit and 4-bit weights plus memory-mapped per-layer embeddings. The same post says LiteRT-LM can process 4,000 input tokens across two distinct skills in under three seconds. Google also says Gemma 4 is available with CPU/GPU support on Android and iOS, accessible through Android’s AICore Developer Preview, and supported on Windows, Linux, macOS via Metal and WebGPU in the browser. On more embedded-class hardware, Google points to Raspberry Pi 5 and Qualcomm’s Dragonwing IQ8, citing 7.6 decode tokens/s on Raspberry Pi 5 CPU and 31 decode tokens/s with IQ8 NPU acceleration. Together, those claims suggest Gemma 4 is not just another open-model release, but part of a broader push toward deployable multimodal and agentic inference on real devices.
