OFFGRID AI OPEN ARENA
RESEARCH PREVIEW

MATCHED MODEL COMPARISON

Same foundation. Different instructions.

E4B PAIR
OffGrid AI OptimizedGemma 4 E4B + OffGrid instructions
vs
Gemma 4 E4BBaseline instruction-tuned model
26B PAIR
OffGrid AI AdvancedGemma 4 26B A4B + OffGrid instructions
vs
Gemma 4 26B A4BBaseline instruction-tuned model
0 / 4,000
Connecting to the pilot...

BEYOND THE ARENA

OffGrid AI ToolKit

Practical AI on a flash drive. Run locally on a compatible Windows or Mac computer, without an internet connection or a subscription for offline use.

Explore the ToolKit

This Arena runs online. Hosting, hardware and runtime differences can affect results on the offline ToolKit.

Privacy & Exports

Data & Privacy

Answers stay in this page until it closes, unless you export them. Remembered statistics contain counts and configuration metadata, not questions or answers. Existing Arena history is kept separate.

Reset statistics

Results Summary

Includes both pairs, ties, losses and sample size. Questions and answers are not included. 


            
        
        
            

                
Grading Report

                
            

            

        
        
            

                
Scoring Methodology

                
            

            

                Each pair uses the same foundation, host and generation
                settings. Only the OffGrid contestant receives the additional
                OffGrid instructions. Each question stands alone. Anonymous
                labels and generation order are shuffled; model identity can
                still be inferred from writing style.
            

            

                GPT-5.2 is the default outside judge. Council is an optional
                experiment with four reviewer seats: two Gemma 4 E4B and two
                Gemma 4 26B. Every reviewer receives only the same grading
                rubric, not the OffGrid contestant instructions. These are two
                underlying models, not four independent judging families. All AI
                judges can have biases and make mistakes.
            

            

                Council uses balanced anonymous labels: each answer occupies
                each A-D position once across the four reviews. All four answers
                and all four reviews must be valid or no winner is recorded.
                Votes are averaged, not summed, keeping totals on the same 0-15
                scale as a single judge. Council and GPT-5.2 results stay
                separate in analytics. No answer is rewritten.
            

            

                Automatic categories use only the question, in a separate
                GPT-5.2 request before answer generation. A selected category
                overrides that step. Categories never enter answer or grading
                prompts. Classification failures are labeled General / Visual
                fallback, not counted as answer failures. Earlier records keep
                their original categories.
            

            

                E4B uses the pinned USB model artifact on a Linux GPU. The 26B
                pair uses hosted OpenRouter inference, not the USB quantized
                file. This measures conditioning within each pair, not overall
                offline runtime equivalence.
            

            
Exact Grader Instructions