Dongwu Securities: Google(GOOGL.US) releases Gemma 4 model capability leap catalyzing the terminal hardware upgrade cycle

Investing in stocks is all about Golden Kylin analyst reports—authoritative, professional, timely, and comprehensive—helping you uncover potential thematic opportunities!

(Source: Zhitong Finance)

Zhitong Finance APP learned that Dongwu Securities issued a research report saying that Google(GOOGL.US) released the Gemma 4 series of open-source models, supporting Agent reasoning, multimodal capabilities (images, video, audio), long context, and multilingual capabilities. The technology focuses on optimizing memory efficiency, lowering the threshold for on-device deployment, and expanding the range of devices covered. With the Apache 2.0 license opening up for commercial use, and combined with deployment in the Android ecosystem, it is expected to drive on-device hardware upgrades and a new round of replacement cycles.

The main points of Dongwu Securities are as follows:

Google releases the Gemma 4 open-source model series, with comprehensive enhancements to Agent and multimodal capabilities

On April 3, Google released its next-generation open-source language model, Gemma 4, including four versions: E2B, E4B, 26B (MoE), and 31B (Dense). The entire Gemma 4 series supports the following capabilities: Agent and complex reasoning: it supports multi-step reasoning and complex logical planning, has autonomous workflow execution capability for Agent scenarios, and can call various tools and APIs. Multimodal: all models natively support image and video processing and perform strongly in tasks such as OCR and chart understanding; in addition, the E2B/E4B versions also support native audio input. Offline code generation: it supports code generation in local environments. Long context: small models support a 128K context window, while large models support up to 256K context, significantly improving the ability to handle long documents and complex tasks. Multilingual capability: it has been trained natively in more than 140 languages.

Technological iteration focuses on memory efficiency optimization and “multimodal capability downscaling,” improving on-device task capacity and expanding device coverage

From the perspective of the technological evolution path, Gemma 4’s iterations optimize core edge-deployment bottlenecks such as memory and interaction capabilities. Specifically: 1) At the model architecture level, it continues the Per-Layer Embeddings (PLE) mechanism. Taking E2B as an example, although it has roughly 5B total parameters, actual inference only needs to load about 2B core weights, while the rest are called on demand by the CPU. This change lowers the hardware usage threshold for end devices, enabling the model to run on existing mid-range devices and expanding the base of on-device AI that can be reached. 2) For long-context capabilities, it uses an “alternating sliding window + global attention” approach and a Shared KVCache design, greatly optimizing memory-use efficiency: most layers only process local tokens, while a small number of layers are responsible for global modeling. Cache reuse avoids repeated computation, reducing KV cache requirements by 74% compared with traditional full attention mechanisms. Against the backdrop of memory constraints on end devices, this optimization directly determines whether models can handle real workloads such as long documents and multi-turn conversations, making it key to turning on-device AI into a productivity tool. 3) At the capability boundaries, for the first time Gemma 4 downscales native multimodal capabilities of vision + audio to 2B-level models, providing a technical foundation for common functions on mobile devices such as understanding screens, voice communication, and cross-application operations. Overall, the firm believes that through architectural innovation, Gemma 4 on the one hand significantly improves the ability of on-device models to handle everyday multimodal tasks, and on the other hand effectively lowers hardware barriers and expands the range of accessible devices, which is of acceleration significance to the pace of the on-device AI industry.

Open-source licensing is fully liberalized, and with the rollout of the Android ecosystem, it drives on-device hardware upgrades and kicks off a new device replacement cycle

From an ecosystem standpoint, earlier generations of the Gemma series used Google’s custom license, which imposed certain limitations on commercial scenarios. This time, Gemma 4 switches to the Apache 2.0 license. Under the absence of mandatory policy constraints, it provides full commercial freedom, significantly reducing the adoption threshold for enterprises and expected to attract more developers and commercial customers back. On the other hand, Gemma 4 will serve as the base model for Gemini Nano 4, and it is planned to be deployed on next-generation flagship Android devices within the year, taking on the role of the next-generation on-device model foundation. According to official disclosures, since its first release, Gemma’s cumulative download volume has exceeded 400 million times, and it has more than 100,000 derivative models, initially forming a Gemmaverse developer ecosystem. The firm believes that under the dual drivers of relaxed open-source licensing and the integration into the Android ecosystem, the on-device model capability upgrades represented by Gemma 4 are likely to significantly expand the boundaries of on-device AI capabilities, further catalyzing upgrades in terminal hardware performance and innovation in new-form product ideas, driving a new cycle of device replacements and breakthroughs across categories.

Risk warning: risks that technological innovation fails to meet expectations, risks of insufficient terminal demand, and risks from the macro environment.

A wealth of information and precise insights—only in the Sina Finance APP