A fresh dive into the One UI 9.5 codebase has uncovered strong hints that Samsung is preparing its next-generation Gauss 4 family of AI models—nearly two years after the company last officially unveiled Gauss 2 back in November 2024. While Samsung has only just kicked off the One UI 9.5 beta program on flagship hardware, strings tucked inside system components point to an ambitious split between lightweight local models and heavier server-backed intelligence.
The model identifiers surface most clearly inside Personal Agent Core—the system service driving One UI 9.5’s proactive, multi-step agentic workflows—as well as Samsung’s new Memory app. Inside Personal Agent Core’s model-selection settings, four Gauss entries stand out alongside two cloud-facing identifiers:
Gauss 4 E2B
Gauss 3.8 3B
GaussA 26B-A4B v0.5
GaussA 30B Latest FP8 DFlash (alongside leftover references to older GaussA 30B builds
v0.6,v0.8, andv0.9)Server models named Galaxy AI Agent v1 and Galaxy AI Bixby Work v1
Even more tellingly, Samsung’s new Memory app explicitly lists Gauss 4 E2B among its supported model IDs and checks the phone’s local storage for a Gauss model file. That suggests on-device inference is front and center for personal context, screen awareness, and quick background actions—themes we have been tracking closely in our coverage of One UI 9.5 agentic AI features and Samsung Memory and One UI 9.5 UX upgrades.
What E2B and 26B-A4B Likely Mean for On-Device Speed
Whenever unfamiliar parameter tags show up in firmware leaks, they usually signal how a model balances reasoning depth against battery drain, memory footprint, and response latency. While Samsung has not published technical documentation for Gauss 4 or GaussA, industry naming conventions around E2B and 26B-A4B give a clear picture of what those labels likely represent.
| Model / Notation | Likely Meaning | Speed & On-Device Trade-Off |
| Gauss 4 E2B | Effective ~2B parameter tier optimized for edge/local execution | Very fast token generation, low RAM pressure, ideal for instant summaries and Memory lookups |
| GaussA 26B-A4B | 26B total parameters with 4B active per token (Mixture-of-Experts style) | Broad knowledge capacity of a 26B model with speed closer to a 4B model |
| GaussA 30B FP8 DFlash | Larger ~30B class using FP8 quantization and fast/flash-style decoding | Higher capability for complex tasks, constrained by memory bandwidth if run locally |
| Galaxy AI Agent v1 / Bixby Work v1 | Server-side agent and enterprise/work assistant endpoints | Offloads heavy multi-app orchestration and large-context work to the cloud |
1. Breaking Down “E2B” in Gauss 4 E2B
In modern edge-AI naming, the “E” in E2B typically stands for Effective parameters (around the 2-billion mark), often paired with techniques like per-layer or selective embeddings that keep active compute lean while preserving useful context. On a phone, a ~2B-class effective footprint is a sweet spot:
Low latency: Small enough to run briskly on a mobile NPU without noticeable lag when summarizing a screen or pulling up a saved memory.
Tighter memory footprint: Leaves breathing room in LPDDR memory alongside One UI 9.5’s visual layers, camera pipelines, and background apps.
Better sustained thermals: Less sustained compute per token helps prevent warm-ups during repeated quick queries.
2. Breaking Down “26B-A4B” in GaussA 26B-A4B
By contrast, 26B-A4B points to a 26-billion total parameter model where only 4 billion parameters (“A4B” = Active 4B) are engaged for a given token—the classic hallmark of a Mixture-of-Experts (MoE) architecture.
Why does that matter for speed? A dense 26B model has to touch huge swathes of weights for every word it outputs, which is slow and memory-heavy on mobile hardware. An MoE layout routes each token to only a small subset of specialized “experts,” letting the model draw on a much larger knowledge base while keeping per-token compute much closer to a 4B-class model. Even then, storing 26B total weights locally still demands serious memory capacity and bandwidth, which may explain why lighter models like Gauss 4 E2B and Gauss 3.8 3B sit right beside it, and why Galaxy AI Agent v1 and Galaxy AI Bixby Work v1 handle heavy server-side lifting.
Why Older GaussA 30B Builds and Memory Checks Matter
Seeing GaussA 30B versions v0.5 through v0.9 alongside a Latest FP8 DFlash tag shows an active engineering pipeline where Samsung has been iterating on quantization and decoding speed rather than settling on a single static checkpoint. Meanwhile, the Memory app checking for a local Gauss model file reinforces how Samsung envisions personal AI working: keeping intimate context—notes, call takeaways, and recent activity—close to the user rather than shipping every snippet off to a server.
As we have also noted in our look at One UI 9.5’s new Galaxy AI contact and memory widget, One UI 9.5 is shaping up to tie together proactive context, smooth UI polish, and deeper assistant actions. Because these references live in unreleased code, final model names, on-device availability, and regional rollouts could still shift before stable release—so it is worth treating these findings as an early window into Samsung’s AI roadmap rather than a locked spec sheet.

















![How to turn on & off Safe Mode on Android [Video] & what can you do in Safe Mode](https://i0.wp.com/nokiapoweruser.com/wp-content/uploads/2021/02/Android-Safe-mode-how-to-video.png?resize=80%2C60&ssl=1)