davidburhans/gevva-e4b

⚡ Gevva e4b: Deep Reasoning AI Decision Engine

<p align="center"> <a href="https://huggingface.co/spaces/davidburhans/gevva-demo"><img src="https://img.shields.io/badge/%F0%9F%A4%97%20Space-Interactive%20Demo-blue.svg" alt="Interactive Demo"></a> <a href="https://pypi.org/project/gevva/"><img src="https://img.shields.io/pypi/v/gevva.svg?logo=pypi&logoColor=white" alt="PyPI"></a> <a href="https://github.com/davidburhans/gevva"><img src="https://img.shields.io/badge/Latency-17.8ms%20(RTX%205090)-orange.svg" alt="Latency"></a> <a href="https://github.com/davidburhans/gevva"><img src="https://img.shields.io/badge/ARC--Challenge-84%25%20Accuracy-brightgreen.svg" alt="ARC-Challenge"></a> <a href="https://opensource.org/licenses/Apache-2.0"><img src="https://img.shields.io/badge/License-Apache--2.0-green.svg" alt="License"></a> </p>

17-millisecond deep reasoning, high-stakes fact checking, multi-tool routing & multi-image verification.

Built on Google's Gemma 4 E4B (4.5B parameters, 42 layers, 128K context) • 84% on ARC-Challenge • 76.6% on JevBench

🤔 What is Gevva e4b? (The 30-Second Explainer)

When you ask ChatGPT or Claude to verify a document or route a customer request, it generates words one token at a time, like a human slowly typing out an explanation. That takes 2 to 5 seconds and burns expensive GPU compute.

That is fine for writing stories, but it is painfully slow and expensive for decisions:

• —"Did the AI hallucinate this medical summary, or is it grounded in the research paper?"
• —"Which of these 10 enterprise APIs should execute this user workflow?"
• —"Did this UI screenshot change in an unexpected way after deployment?"
• —"Did the student's solution satisfy all 5 steps in the grading rubric?"

The Solution: An Instant "Reflex Engine" with Deep Reasoning

Psychologist Daniel Kahneman famously described human thought in two modes:

• —System 1 (Fast Reflexes): Instant decisions in 15–20 milliseconds (dodging a ball, recognizing a face).
• —System 2 (Slow Reasoning): Writing essays or working out long equations step-by-step.

Gevva e4b is the 4.5B deep reasoning model in the Gevva family. In a single 17.8 millisecond forward pass, it evaluates evidence and outputs confident, mathematically calibrated decision probabilities.

┌────────────────────────────────────────┐       ┌────────────────────────────────────────┐
│         SYSTEM 1: GEVVA e4b            │       │       SYSTEM 2: CHATGPT / CLAUDE       │
│  ⚡ Evaluates decisions in 17.8 ms     │  vs   │  🐢 Generates text word-by-word        │
│  🎯 84% ARC-Challenge accuracy        │       │  ⏳ Takes 2,000 - 5,000 milliseconds   │
│  💰 90%+ cheaper compute cost          │       │  💸 Expensive GPU server bills         │
│  🔍 Best for: Fact-checks, routing,    │       │  ✍️ Best for: Creative writing, long   │
│     guardrails, and visual audit       │       │     essays, and coding from scratch    │
└────────────────────────────────────────┘       └────────────────────────────────────────┘

🚀 3-Line Quickstart

from gevva import load

# 1. Load the decision engine (GPU or CPU)
engine = load("davidburhans/gevva-e4b")

# 2. Instant Fact Check / Hallucination Detection
verdict, probs = engine.grade(
    premise="Albert Einstein won the Nobel Prize in Physics in 1921 for his discovery of the photoelectric effect.",
    hypothesis="Einstein won the Nobel Prize for his theory of General Relativity."
)

print(verdict)
# Output: 'contradiction' (p_contradiction = 0.96)

Smart Tool & API Routing

tools = [
    "send_email(to, subject, body): Sends an email message to a contact",
    "search_database(query): Queries internal company records and docs",
    "refund_charge(charge_id, reason): Issues a payment refund to a customer",
]

best_idx, scores = engine.rerank(
    premise="A customer says they were double-billed and wants their money back.",
    options=tools
)

print(f"Selected: {tools[best_idx]}")
# Output: Selected: refund_charge(...)

📊 Evaluated Benchmarks on RTX 5090

Gevva e4b was evaluated on standard reasoning batteries on an NVIDIA GeForce RTX 5090:

🛠️ Architecture Details

• —Foundation Backbone: Google's gemma-4-E4B-it (4.5B parameters, 42 layers).
• —Classification Head: Last-token pooled representation normalized with Gemma4RMSNorm into a 3-class linear head.
• —Label Convention:
• —0: Contradiction / Refutes / Invalid
• —1: Entailment / Supports / Valid
• —2: Neutral / Unverifiable / Inconclusive
• —Calibration: Trained with strictly proper Brier score calibration ($ECE = 0.0224$ on validation).

📄 License & Citations

Released under the Apache 2.0 License.

@software{gevva2026,
  author = {David Burhans},
  title = {Gevva: Multimodal 128K System 1 Decision Engine},
  year = {2026},
  url = {https://github.com/davidburhans/gevva}
}