Gemma 4 Release: On-Device Models from E2B to 31B
April 02, 2026 · 5 min read
Gemma 4 is Google’s latest open-weight family, announced April 2, 2026: four checkpoints aimed at everything from ultra-mobile browsers to serious local and self-hosted setups. The lineup pairs 2B and 4B effective “E” models (built with the same Per-Layer Embedding ideas that made Gemma 3n practical on phones) with a 31B dense flagship and a 26B mixture-of-experts variant that activates 4B parameters per token. Smaller models carry a 128K context window; the larger two expand to 256K. Gemma 4 is licensed under Apache 2.0—a shift from the Gemma Terms of Use that applied to earlier generations—so downstream use, modification, and redistribution follow a familiar open-source playbook.
Weights are on Hugging Face and Kaggle. For full technical detail, see Google’s Gemma 4 model card and core overview.
Enclave AI and Gemma 4
Enclave AI does not ship Gemma 4 today. The current App Store build still focuses on the Gemma 3 and Gemma 3n lineups and our other supported models. Gemma 4 support is in progress and we expect to roll it out in the next few days—watch for an app update, then pick Gemma 4 from the model list once it appears.
What shipped: the four sizes
| Variant | Role | Notes |
|---|---|---|
| Gemma 4 E2B | Edge, phone, browser | “Effective” ~2B parameters; PLE-heavy architecture tuned for constrained memory |
| Gemma 4 E4B | Stronger edge / tablet | ~4B effective; same mobile-first design philosophy as E2B |
| Gemma 4 26B A4B | Throughput + reasoning | MoE: 4B active per forward pass; full 26B weights still need to be resident for routing (see below) |
| Gemma 4 31B | Maximum quality in the family | Dense model bridging server-grade behavior and ambitious local setups |
Across the family, Google emphasizes reasoning (including configurable thinking modes), coding, built-in function calling, and native system-prompt / system-role support—useful for agents and structured assistants without brittle prompt hacks.
Multimodal and long context
Gemma 4 is multimodal by design: text and images (with variable aspect ratio and resolution) on all sizes; video and audio are part of the stack, with audio emphasized on the E2B and E4B models for on-device scenarios.
For local AI users, the headline is not just benchmarks but usable context: 128K tokens on the small models and 256K on 31B and 26B A4B. Long context still costs KV-cache memory on top of static weights—plan headroom if you actually fill the window. Our KV cache explainer walks through why long chats and big documents scale memory the way they do.
Memory: what Google quotes for inference
Google publishes approximate load-only weight memory (not full runtime with max context). For Q4_0 (4-bit) inference:
| Model | Approx. Q4_0 weight memory |
|---|---|
| Gemma 4 E2B | ~3.2 GB |
| Gemma 4 E4B | ~5.0 GB |
| Gemma 4 26B A4B | ~15.6 GB |
| Gemma 4 31B | ~17.4 GB |
BF16 numbers are higher (see the official table); SFP8 sits between BF16 and Q4_0.
Important MoE caveat: For 26B A4B, Google explicitly notes that although only 4B parameters are active per token, all 26B parameters must be loaded to keep expert routing and inference fast. In practice, that means disk and RAM pressure look like a large model, not like a 4B dense model—unlike some MoE deployments where only active experts stay hot. If you are sizing a Mac or a home server, budget for the full checkpoint, then add context and framework overhead.
What Gemma 4 means for iPhone and Mac
-
Phones and tablets: The E2B and E4B targets align with the same story as Gemma 3 and Gemma 3n on iOS—models engineered so that reasonable quantizations can stay within unified memory on recent Apple silicon devices. Once Enclave adds Gemma 4, the same tradeoffs apply; our quantization guide for iPhone and Mac is still the right reference for bit depth and real-world RAM use.
-
Mac desktops and workstations: 31B at Q4 is already in the ~17 GB static weight ballpark before context—feasible on 24 GB+ machines with headroom, tighter on 16 GB. The 26B MoE is not a free lunch for memory despite the 4B activation count.
-
Privacy: Nothing about Gemma 4 changes the core tradeoff: open weights + local inference means your prompts and documents never need to touch a cloud API. That is the same promise we care about across offline AI and local-first workflows.
How this fits next to Gemma 3 and the wider field
Gemma 3 established Google’s push into efficient, on-device-capable open models with strong text performance and large context. Gemma 3n doubled down on mobile-first design. Gemma 4 is the generational bump: longer context on the big tiers, richer multimodal surface (including video and audio where advertised), agent-oriented features (function calling, system role), and Apache 2.0 for simpler product integration.
Compared with other 2026 families—such as Qwen 3.5’s eight-model spread—Gemma 4 is a tighter, four-model line with a very clear split between E-series edge and 26B/31B desktop or server. Neither approach is “better”; they optimize for different deployment ladders.
Bottom line
Gemma 4 refines Google’s bet that open weights and device-local inference belong in the same sentence as frontier-class behavior—especially for reasoning, tools, and long documents. The E2B/E4B pair is the one to watch for private AI on phones; 31B is the quality anchor for serious local chat; 26B A4B is a throughput-oriented MoE as long as you respect full-model memory. In Enclave, Gemma 4 will slot into that same privacy-first story as soon as the upcoming update lands—until then, Gemma 3n remains the cutting edge of Google’s stack inside the app.