NEW GEMMA 4 beats GPT-5.4: The A4B Model

2026-04-03 · Source: Discover AI · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Data Science & Analytics, Robotics & Autonomous Systems · Depth: Expert, extended

Summary

Google has released the Gemma 4 series of open-source models under an Apache 2 license, including 2B, 4B, 26B Mixture-of-Experts (MoE), and 31B dense models. This analysis focuses on live testing the 26B MoE (which activates 3.88B parameters) and the 31B dense model using a complex "elevator puzzle" designed to assess causal reasoning and logical problem-solving without external tools. The 4B active MoE model consistently demonstrated superior self-reflection, strategic planning, and constraint adherence, ultimately finding a valid 10-button press solution. In contrast, the 31B dense model struggled with optimization, often getting stuck in local minima and violating puzzle constraints, leading to invalid or suboptimal solutions. The 4B MoE's performance rivaled or exceeded larger proprietary models like GPT-5.4 (non-X-High) on this specific task.

Key takeaway

For AI Scientists and Machine Learning Engineers evaluating open-source LLMs for complex logical tasks, you should prioritize the Gemma 4 26B MoE (4B active) model. Its demonstrated self-correction and strategic planning capabilities make it a strong contender for applications requiring robust causal reasoning, potentially outperforming larger dense models and even some proprietary alternatives on such challenges. Consider its 31B dense counterpart primarily as a foundation for extensive fine-tuning.

Key insights

Gemma 4's 4B active MoE model excels in complex logical reasoning and self-correction, outperforming its 31B dense counterpart.

Principles

• Smaller MoE models can surpass larger dense models in complex reasoning.
• Self-reflection and iterative checking are crucial for robust problem-solving.
• Pure intelligence testing should exclude external agentic tools.

Method

The "elevator puzzle" assesses LLM causal reasoning by requiring shortest path optimization under complex mathematical button functions, energy limits, and floor caps, without external solvers.

In practice

• Prioritize Gemma 4 4B MoE for tasks requiring deep logical reasoning.
• Use Gemma 4 31B as a base for fine-tuning specific domain tasks.
• Design evaluation puzzles with complex constraints to reveal model intelligence.

Topics

• Gemma 4 Models
• Mixture-of-Experts
• Complex Logic Puzzles
• Causal Reasoning
• LLM Performance Benchmarking

Best for: AI Scientist, Machine Learning Engineer, Director of AI/ML

Related on AIssential

• Gemma 4 31B vs Qwen3.5 27B: Inference Speed, Token-Efficiency, Accuracy, and Memory Consumption — Gemma 4 31B generally outperforms Qwen3.5 27B in accuracy and consistency, with better generalization.
• Introducing Gemma 4 models on Amazon Bedrock — Gemma 4 models on Amazon Bedrock offer flexible, open-weight AI with multimodal capabilities and managed inference for diverse workloads.
• Gemma 4 31B and 26B A4B: Architecture and Memory Consumption — New Gemma 4 and Trinity-Large-Thinking models offer diverse architectures and memory footprints for large language model deployment.
• The Sequence AI of the Week #839: Gemma 4 and the Compression of Intelligence — Gemma 4 transforms frontier AI capabilities into practical, deployable infrastructure for diverse applications.
• Google’s Gemma 4: Is it the Best Open-Source Model of 2026? — Google's Gemma 4 family offers versatile, open-weight LLMs optimized for diverse environments, from edge devices to consumer GPUs.
• Gemma 4 12B MTP Local Test | Coding, OCR, Visual RAG with llama.cpp — The Gemma 4 12B MTP model offers multimodal capabilities and efficiency, nearing larger MoE models, but with performance trade-offs.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by Discover AI.