Casual conversations
Daily chat, natural tone, fast replies. Ideal for general assistance without burning the “big brain” on every question.
Junior series · smallest in Xeretron
Xeretron Junior E4B
The smallest model in the family. Built on Gemma 4 E4B (google/gemma-4-E4B-it). Perfect for casual conversations, document reading, and small agents. It is the base model to share real AI across the company at low cost: handles email and connects with XMail to reply to mail, manage the calendar, and more. Up to 290 t/s (depending on GPU).
E4B = ~4.5B effective parameters (~8B with embeddings / PLE). Specs from the Gemma 4 model card. Junior is the Xeretron fine-tune on that base for everyday enterprise use.
Junior highlights
Small in size. Big in coverage.
It does not need a full rack: enough for the whole company to talk to real AI.
Document reading
PDFs, screens, OCR, and visual understanding (Gemma 4 multimodal). Summarize, extract, and answer questions about work documents.
Small agents
Native function calling for micro-agents: tickets, FAQs, short flows, and panel tools — without orchestrating a Frontier fleet.
AI for the whole company
The base model to share real intelligence at low cost: more concurrent users, less VRAM, same Xeretron node.
Email with XMail
Connects with XMail to reply to mail in your voice, prioritize urgent items, and keep your inbox off third-party clouds.
Calendar under control
With XMail: plan the day, remember appointments, and manage the calendar from the same assistant that handles email.
Base · Gemma 4 E4B
Edge-ready. Multimodal. Apache 2.0.
Open weights via google/gemma-4-E4B-it (Google DeepMind). Built for laptop and light server: native text, image, and audio, 128K context, configurable reasoning.
~4.5B effective · PLE
Per-Layer Embeddings: large tables for fast lookups; active compute stays small. More intelligence per watt on the node.
Text · image · audio
Documents, screens, charts, and audio (ASR / translation). Ideal for inbox and everyday attachments.
Tools & agents
Native function calling + system prompt: Junior plugs into XMail, light MCPs, and short enterprise flows.
Apache 2.0
Clean commercial use: fine-tune, redistribution, and node deployment with NOTICE — without legacy Gemma Terms.
Base reference
Public Gemma 4 E4B numbers.
Figures from the model card (instruction-tuned). Junior inherits this base and specializes it for the Xeretron + XMail ecosystem.
| Benchmark | Gemma 3 27B | Gemma 4 E4B (base Junior) |
|---|---|---|
| MMLU Pro | 67.6% | 69.4% |
| LiveCodeBench v6 | 29.1% | 52.0% |
| GPQA Diamond | 42.4% | 58.6% |
| Tau2 (avg) | 16.2% | 42.2% |
| MMMU Pro | 49.7% | 52.6% |
| MATH-Vision | 46.0% | 59.5% |
| OmniDocBench 1.5 ↓ | 0.365 | 0.181 |
Source: model card Gemma 4 E4B-it / Google DeepMind. ↓ = lower is better. Junior does not replace evaluation on your hardware.
Specifications
Junior technical sheet.
| Attribute | Detail |
|---|---|
| Name | Xeretron Junior E4B |
| Base | Gemma 4 E4B-it · google/gemma-4-E4B-it |
| Parameters | ~4.5B effective · ~8B with embeddings (PLE) · 42 layers |
| Context | 128K tokens |
| Modalities | Text · image · audio · tool / agents |
| Xeretron specialty | Enterprise at low cost · chat · docs · XMail · calendar · small agents |
| Throughput | Up to ~290 t/s (depending on GPU) |
| Base license | Apache 2.0 — LICENSE / NOTICE |
License · Apache 2.0
On your node
Fits where the big ones do not.
Junior is built to run light: more seats, more conversations, less GPU. Pair it with XMail and save Omega/Titan/Developer for when the work demands it.
- Ideal as the enterprise default in the panel.
- XMail: mail + calendar on the same node.
- OpenAI-compatible API for internal apps.
- Scale users without multiplying hardware.
Recommended stack
- Consumer or edge GPU · light quantization
- Junior + XMail profile in the panel
- Small agents / FAQs / inbox
- Step up to Developer/Turbo/Titan/Omega when needed
Next step
Put real AI on every desktop.
Junior for day-to-day and XMail. The big brains when the problem deserves them.