FlashPapers
Google DeepMind Technical Report

Gemini: A Family of Highly Capable Multimodal Models

By the Gemini Team, Google • Published December 2023 (v5 May 2025)

TL;DR: Google introduces Gemini 1.0, a family of natively multimodal models (Ultra, Pro, Nano) trained jointly across text, images, audio, video, and code. Gemini Ultra is the first model to surpass human-expert performance on the MMLU benchmark (90.04%) and advances the state of the art in 30 of 32 major benchmarks.

Natively Multimodal MMLU > 90% Ultra / Pro / Nano Joint Pre-training

Why This Matters

Historically, building AI systems that can see, hear, and read required stitching together completely separate models. An image encoder was glued to a text generator, which was then patched to an audio transcriber. While functional, this "cascade" approach is highly fragile. Essential context is lost at the boundaries between models; for example, the subtle emotional tone in an audio waveform or the precise spatial layout of a complex chart gets flattened when translated into intermediate text representations.

Gemini changes this paradigm by being natively multimodal from day one. By training a single, unified architecture directly on interleaved sequences of text, images, audio, and video, the model learns to build a holistic representation of the world. It can reason seamlessly across different modalities, enabling educational, scientific, and creative workflows that were previously impossible.

The Big Idea

"We trained Gemini models jointly across image, audio, video, and text data for the purpose of building a model with both strong generalist capabilities across modalities alongside cutting-edge understanding and reasoning performance in each respective domain."

Instead of processing visual or auditory inputs into text before feeding them to a transformer, Gemini maps all inputs directly into a shared token space. This allows the core decoder-only Transformer to process visual, auditory, and textual sequences in exactly the same way. The models are scaled across three distinct sizes: Ultra (for highly complex reasoning), Pro (for optimized cost and latency), and Nano (highly distilled 1.8B and 3.25B models optimized for on-device execution).

How It Works

1. Interleaved Token Sequences

In Gemini, different modalities are converted into sequences of tokens. Images are split into patches or encoded as discrete visual tokens; audio signals at 16kHz are converted into dense feature representations; video is treated as a sequence of frames interleaved naturally within the large 32k context window.

Interactive Demo: Native Multimodal Sequencer

Construct a custom multimodal prompt and observe how Gemini packages it into a single sequence for the Transformer.

Add Modality Blocks
Interleaved Token Sequence (Input Stream)

Click blocks on the left to interleave inputs...

Transformer Decoder Output: Waiting for sequence...
"I'm ready to receive your natively interleaved sequence."

2. Training Infrastructure & Goodput

Training Gemini Ultra required a massive fleet of TPUv4 accelerators deployed in "SuperPods" of 4096 chips, connected via dynamic optical switches. At this scale, hardware failures are frequent. Traditional checkpointing to persistent storage is too slow, dropping goodput (useful training time).

To solve this, Google utilized redundant in-memory model state replicas. When a hardware node fails, the system recovers directly from an intact in-memory replica, boosting training goodput from 85% to 97%.

Interactive Simulation: Checkpoint vs. In-Memory Recovery

Simulate a hardware failure during training to see how in-memory replication preserves training efficiency.

Simulation Parameters
Frequent Failures Rare Failures
Traditional Disk Checkpoint (85% Goodput) Idle
Gemini In-Memory Replication (97% Goodput) Idle
System Log:
Simulation ready. Click "Start Training Run".

3. Uncertainty-Routed Chain-of-Thought

To achieve its record-breaking 90.04% MMLU score, Gemini Ultra utilizes a novel technique called Uncertainty-Routed Chain-of-Thought (CoT).

Rather than always relying on expensive CoT reasoning (which can sometimes degrade performance if the model becomes inconsistent), the system first evaluates the problem's difficulty. If the model's confidence is above a certain threshold, it answers greedily. If the confidence is low, it routes the query to generate $k$ chain-of-thought samples and takes a consensus vote.

Interactive Tool: Uncertainty Router

Configure the routing threshold and confidence to see how the model dynamically switches between greedy and CoT paths.

Router Configuration
Routing Decision Flow
Input Prompt
Uncertainty Router
Greedy Decoding
CoT Consensus ($k$ samples)
Expected Accuracy: On MMLU, this routing strategy achieves 90.04%, outperforming pure CoT@32 (85.0%) and pure greedy decoding (84.0%).

Results

Gemini Ultra sets a new state of the art across almost all major benchmarks, outperforming GPT-4 on key text reasoning and mathematical evaluations, and showing massive gains in multimodal benchmarks without task-specific tuning.

Benchmark Comparison: Gemini Ultra vs. Competitors

Benchmark Domain Gemini Ultra GPT-4 / GPT-4V PaLM 2-L

Limitations & Open Questions

Despite its state-of-the-art capabilities, the Gemini family still faces core challenges inherent to modern LLMs:

  • Hallucinations: Like other language models, Gemini can generate plausible-sounding but factually incorrect outputs. Google addresses this during post-training SFT and RLHF, but the issue is not fully resolved.
  • High-Level Reasoning: The paper explicitly notes that Gemini still struggles with tasks requiring deep causal understanding, logical deduction, and counterfactual reasoning.
  • Socioeconomic Bias in Vision: Evaluation on the Dollarstreet dataset reveals that image understanding is less effective for images sourced from lower socioeconomic regions compared to North America and Europe.

Glossary

Native Multimodality

An architecture trained on multiple modalities (text, images, audio, video) simultaneously from the start, rather than using separate pre-trained encoders stitched together.

Goodput

The ratio of time spent computing useful new gradient steps over the total elapsed time of the training job. High goodput indicates minimal overhead from failures and recovery.

MMLU (Massive Multitask Language Understanding)

A holistic benchmark testing knowledge and reasoning across 57 subjects spanning STEM, humanities, and social sciences.

Silent Data Corruption (SDC)

Rare hardware faults that cause incorrect mathematical computations without throwing errors. Gemini mitigated this using deterministic replay and proactive scanners.

Citation

@article{geminiteam2023gemini,
  title={Gemini: A Family of Highly Capable Multimodal Models},
  author={Gemini Team, Google},
  journal={arXiv preprint arXiv:2312.11805},
  year={2023}
}
Made with Flash Papers — make your own