Google Gemini: The Gemini era begins

The launch and the specific claim

On 6 December 2023, Google introduced Gemini, described by Sundar Pichai and Demis Hassabis as the company's "largest and most capable AI model." The announcement, published on the official Google blog, made three concrete claims. First, Gemini 1.0 was released in three sizes (Ultra, Pro, and Nano). Second, it was "built from the ground up to be multimodal," meaning it was trained on text, code, audio, image, and video rather than assembled from separately trained components. Third, Gemini Ultra scored 90.0% on the MMLU knowledge benchmark, which Google presented as the first time a model had reportedly surpassed a human-expert baseline on that test. The same accompanying Gemini technical report documents the architecture, the evaluation protocols, and the caveats those headline numbers carry.

The historically useful distinction here is between the narrow, documented event and the broad judgment placed on it. The narrow event is a product-and-research announcement with specific variants, benchmarks, and rollout dates. The broad judgment is that Google consolidated a scattered set of AI efforts into a single frontier model family and rebuilt its consumer and developer offerings around it. Both are true, but they answer different questions, and conflating them tends to inflate a single launch into an inevitability it was not.

Why one model family: the merger and the ChatGPT shock

Gemini did not appear in a vacuum. Its most important predecessor was organizational. In April 2023, Pichai announced Google DeepMind, merging the Google Brain team from Google Research with DeepMind into a single unit led by Hassabis, with Jeff Dean as chief scientist. For years those two groups had pursued overlapping but distinct research agendas. Gemini was the first frontier model produced by the combined organization, and the name (the twins) reads naturally as a marker of that union.

The competitive pressure was equally concrete. OpenAI's ChatGPT arrived in November 2022, and Google's first public conversational response came on 6 February 2023 with Bard, an experimental service powered by LaMDA. In May 2023 Google introduced PaLM 2, an upgraded language model that took over as Bard's engine and powered more than two dozen products. The constraint Gemini addressed was therefore both technical and strategic: Google had capable models (LaMDA, PaLM, PaLM 2) but no single, natively multimodal frontier model designed from the outset to span consumer products, developer APIs, and on-device use. Gemini was the attempt to unify that direction.

Natively multimodal, and why that framing mattered

The most substantive technical claim was that Gemini was multimodal by construction rather than by retrofit. A common earlier pattern was to train a strong text model and then bolt on a separate vision encoder, stitching modalities together after the fact. Google's technical report describes Gemini instead as trained jointly across modalities from the start, so that a single model could accept interleaved sequences of text, images, audio, and video and reason across them.

The practical significance of this is best stated carefully. Native multimodal training does not guarantee superior performance on any given task, and it does not make modalities interchangeable: an image patch, an audio waveform, and a text token carry different structure, and the model still learns modality-specific behaviour. What the design changes is the feasible experiment. A model that ingests a diagram, the audio of someone describing it, and a written question in one pass can, in principle, relate those inputs without a brittle handoff between separate systems. Whether it does so reliably is an empirical question the benchmarks were meant to probe, not something the architecture settles on its own.

Three sizes for three deployment surfaces

Gemini 1.0 shipped as a deliberately tiered family, and the tiers mapped onto where the model would run. Ultra was the largest, aimed at highly complex tasks. Pro was the general workhorse "for scaling across a wide range of tasks," and it became the model behind a fine-tuned version of Bard on launch day and the model exposed to developers through the API from 13 December 2023. Nano was the efficiency tier, small enough to run on a device. According to the technical report, Nano came in two distilled variants of roughly 1.8 billion and 3.25 billion parameters, and Nano powered on-device features such as Recorder summarization and Smart Reply on the Pixel 8 Pro.

This tiering is worth dwelling on because it reflects a real engineering discipline rather than marketing. Distilling a small model that runs locally and a large model that runs in a datacentre from a shared design lets a company amortize research across very different cost and latency budgets. It also means "Gemini" from day one referred to a spectrum of capability, not one artifact. A reader comparing "Gemini" against a rival model should always ask which variant, because the gap between Nano and Ultra is large and intentional.

What the benchmarks actually said, and the CoT@32 asterisk

The headline figure was 90.0% on MMLU, the Massive Multitask Language Understanding benchmark that spans 57 subjects from mathematics to law, introduced by Dan Hendrycks and colleagues in 2020. Google reported that Gemini Ultra was the first model to exceed the roughly 89.8% human-expert baseline the report used. On the multimodal side, Gemini Ultra scored 59.4% on MMMU, a college-level multi-discipline reasoning benchmark released weeks earlier, ahead of the contemporaneous GPT-4V result near 56%.

The honest caveat sits inside the same report. The 90.0% MMLU figure was obtained with an elaborate prompting strategy, chain-of-thought with 32 samples (reported as CoT@32), not a standard few-shot evaluation. On an apples-to-apples 5-shot comparison, the technical report shows Gemini Ultra at 83.7%, below the 86.4% that OpenAI's GPT-4 Technical Report had reported for the same 5-shot setting. In other words, the record depended on the measurement protocol, and under the protocol its main rival was quoted on, Gemini Ultra did not lead. This is not evidence of dishonesty; it is a reminder that a benchmark number is a task-specific result under stated conditions, never a measure of general intelligence, and that comparisons are only meaningful when the evaluation method is held fixed.

The first independent check came within two weeks, and it tested the variant people could actually use. Because Ultra was not yet available, a team led by researchers at Carnegie Mellon University compared Gemini Pro with OpenAI's GPT-3.5 Turbo and GPT-4 Turbo and with the open Mixtral model, and released its code and data so others could reproduce the runs. Their paper, An In-depth Look at Gemini's Language Abilities, found Gemini Pro close to but slightly behind GPT-3.5 Turbo on every task they benchmarked. On MMLU in the 5-shot setting it scored 65.2%, against 67.8% for GPT-3.5 Turbo and 80.5% for GPT-4 Turbo. The authors' MMLU numbers ran slightly below Google's own, a difference they attributed to sensitivity to prompts, and their code-generation result for Gemini Pro fell well short of the figure Google reported, which they suspected reflected different prompting. They also catalogued concrete failure modes, among them errors in arithmetic with many-digit numbers, sensitivity to the order of multiple-choice answers, and aggressive content filtering, alongside comparatively strong results on generation in languages other than English and on longer reasoning chains. Seen from outside, the lesson matches the CoT@32 asterisk: the Gemini that developers could call in December 2023 was a GPT-3.5-class model, and the variant behind the record could not yet be checked by anyone outside Google.

The competitive field of late 2023

Gemini entered a crowded frontier. GPT-4, released in March 2023, was already a large-scale multimodal system accepting image and text input and producing text, and OpenAI had extended it with vision (GPT-4V) and, in late 2023, image generation and voice. Anthropic had shipped Claude 2 with a long context window aimed at document work, and Meta had released Llama 2 in July 2023 as openly available weights, seeding a large ecosystem of fine-tuned open models. The competing approaches differed on more than raw capability: closed frontier APIs (OpenAI, Anthropic, Google) versus open-weight distribution (Meta), and monolithic dense models versus the mixture-of-experts designs that would soon become common.

Placed in that field, Gemini's distinctive bets were native multimodality across the whole family and vertical integration with Google's distribution (Search, Android, Pixel, Workspace, Cloud). It was not the first multimodal frontier model, and it did not win every benchmark. Its strategic novelty was the breadth of surfaces a single model family was meant to serve, from a phone's on-device assistant to an enterprise API.

The staggered rollout and a demo that overpromised

A recurring source of confusion is that "Gemini launched" on 6 December 2023 does not mean everything launched at once. On that date, Bard gained a fine-tuned Gemini Pro, and Pro reached developers the following week. Gemini Ultra, the variant behind the record benchmark, was not generally available. It arrived over two months later, on 8 February 2024, when Google renamed Bard to Gemini and launched Gemini Advanced with Ultra 1.0 inside the Google One AI Premium plan at $19.99 per month, alongside new Android and iOS apps. Product availability differed by variant, region, and date, and the launch-day demonstrations should not be read as unrestricted live functionality for every user.

The clearest illustration of that gap was the "Hands-on with Gemini" video. Google's own developer post, How it's Made: Interacting with Gemini through multimodal prompting, documents that the interactions were produced by prompting Gemini with still image frames and text, then editing the sequence for a smooth presentation. The polished video implied a fluid, real-time spoken conversation that the model was not, at that moment, performing live. The underlying capabilities were real, but the presentation compressed latency and staged the inputs, and the episode became a widely cited caution about the distance between a launch reel and a shipping product.

The infrastructure underneath: TPUs

Frontier models are inseparable from the hardware that trains them, and Gemini was a Google-silicon story. The technical report states that Gemini was trained on Google's Tensor Processing Units (the v4 and v5e generations), and Google's cloud announcement the next day confirmed that Gemini "was trained on, and is served, using TPUs". That same post introduced Cloud TPU v5p, described as up to 2.8 times faster than v4 at training large language models and roughly four times more scalable per pod.

This detail matters for placing the event accurately. Part of what let Google mount a credible frontier effort was that it designs and operates its own accelerators at scale, rather than competing purely for third-party GPUs. The model, the training system, and the serving fleet were co-designed, which is a structural advantage distinct from any single benchmark. It also complicates simple capability comparisons, because efficiency, cost, and serving throughput depend on the hardware and systems layer as much as on the model diagram.

What came next: 1.5, long context, and the rebrand

Gemini's most consequential downstream influence was fast and internal to Google. Just one week after Gemini Advanced, on 15 February 2024, Google announced Gemini 1.5. Gemini 1.5 Pro used a mixture-of-experts architecture, activating only the most relevant expert sub-networks per input, and it delivered quality Google described as comparable to 1.0 Ultra while using less compute. Its signature feature was context length: a standard 128,000-token window and an experimental 1 million-token window, enough to process about an hour of video, eleven hours of audio, or codebases exceeding 30,000 lines in a single prompt. Long context became a defining axis of competition across the industry in 2024.

The other durable consequence was branding and consolidation. "Gemini" replaced "Bard" as the name of Google's assistant, and over the following years the family expanded into successive generations. Google DeepMind's current Gemini overview page shows how far that consolidation went: by 2026 the brand spans many specialized variants, from efficient Flash models to large Pro and deep-reasoning tiers, plus audio, robotics, and image-generation offshoots. The December 2023 launch is best understood as the moment that brand and that unified model direction began, not as the moment its capabilities peaked.

How to place the Gemini launch in AI history

The evidence supports a measured reading. The narrow, documented event is that on 6 December 2023 Google DeepMind released a natively multimodal model family in three sizes, reported a leading MMLU score under a specific prompting protocol, wired Gemini Pro into Bard and its developer API immediately, and shipped the top Ultra tier and the Gemini rebrand two months later. The broad historical judgment is that this launch marked Google's reorganization of its entire AI product and research direction around one model family, backed by its own TPU infrastructure, in direct response to a competitive frontier defined by OpenAI, Anthropic, and Meta.

What Gemini did not do is equally important to state. It did not win every benchmark, its record number carried a protocol asterisk, and its most striking demo was staged. None of that erases its significance; it clarifies it. A useful way to navigate outward from this event is to hold three questions apart. Which earlier decision made it possible (the Brain and DeepMind merger, the PaLM lineage, and Google's silicon)? Which specific claim is documented and under what conditions (the tiered family, the CoT@32 MMLU result, native multimodality)? And which later work actually delivered on the promise (Gemini 1.5's long context, and the successive generations that followed)? Keeping those levels distinct turns a launch reel into a legible piece of history, one where a company's frontier ambitions and its real, staggered, benchmark-qualified capabilities can both be seen clearly.

References

• Pichai and Hassabis (2023). Introducing Gemini: our largest and most capable AI model.
• Gemini Team, Google (2023). Gemini: A Family of Highly Capable Multimodal Models.
• Pichai (2023). Announcing Google DeepMind.
• Google (2023). An important next step on our AI journey.
• Google (2023). Introducing PaLM 2.
• OpenAI (2023). GPT-4 Technical Report.
• Hendrycks et al. (2020). Measuring Massive Multitask Language Understanding.
• Yue et al. (2023). MMMU: A Massive Multi-discipline Multimodal Understanding and Reasoning Benchmark for Expert AGI.
• Google Cloud (2023). Enabling next-generation AI workloads: Announcing TPU v5p and AI Hypercomputer.
• Google for Developers (2023). How it's Made: Interacting with Gemini through multimodal prompting.
• Google (2024). Bard becomes Gemini: try Ultra 1.0 and a new mobile app today.
• Pichai and Hassabis (2024). Our next-generation model: Gemini 1.5.
• Google DeepMind (2024). Gemini.
• Akter et al. (2023). An In-depth Look at Gemini's Language Abilities.