Gemini 3 Pro Preview High

Google explicitly identifies Gemini 3 Pro as a sparse Mixture-of-Experts (MoE) transformer-based model, a shift from the dense architecture described in some preview marketing. While the model card cites foundational MoE research (e.g., Shazeer et al., 2017; Fedus et al., 2021), it lacks specific details on the number of experts, routing mechanisms, or the exact architectural modifications that enable its 1-million-token context window. The documentation confirms it is a 'native multimodal' model rather than a modular system, but the specific integration of modality-specific encoders remains high-level.

The training data is described in broad categories: web documents, books, code, images, audio, and video. While Google provides some high-level estimates for the 'Pro' family (e.g., ~3T text tokens, 1B image-text pairs), it does not provide a specific percentage breakdown or detailed filtering/cleaning methodology for the Gemini 3 Pro Preview High variant. The use of 'publicly available' and 'licensed' data is mentioned without naming specific sources, and the inclusion of synthetic data is acknowledged but not quantified.

The model uses a SentencePiece unigram tokenizer with a vocabulary size of 256,000 tokens, consistent across the Gemini family. This tokenizer is publicly accessible via the Google 'vertexai' and 'generative-ai' Python SDKs, allowing for local verification of token counts and normalization behavior. Documentation confirms it supports unified processing across text, code, and multimodal transcripts, though the specific 'thinking tokens' used in the High reasoning mode are hidden from the final API output.

Total and active parameter counts for Gemini 3 Pro are not officially disclosed. While third-party analysis from Artificial Analysis suggests the model is significantly larger than its predecessors due to its factual recall performance, Google maintains a policy of not releasing these figures for its frontier models. As a sparse MoE model, the lack of information regarding the number of experts or active parameters per token represents a major transparency gap.

Google confirms the model was trained on TPU v5p/v6 infrastructure but provides no data on total GPU/TPU hours, energy consumption, or carbon footprint. There are no public estimates of the training cost or duration. The documentation focuses on the scalability of TPU Pods rather than the specific resources consumed by this model version.

Google provides scores for several standard benchmarks (GPQA Diamond: 94.3%, ARC-AGI-2: 77.1%, SWE-Bench Verified: 80.6%). However, the evaluation code and exact prompts used for these internal 'verified' scores are not fully public. While some results are cross-referenced on leaderboards like the ARC Prize, the 'Preview High' variant's specific 'thinking' depth makes exact reproduction difficult for external auditors without access to the same internal configuration.

The model consistently identifies as Gemini 3 Pro and is aware of its versioning (e.g., distinguishing between 3.0 and 3.1 in API responses). It accurately describes its multimodal capabilities and the 'thinking' parameter. There are no documented cases of the model claiming to be a competitor's product or denying its nature as a Google-developed AI.

The model is released under a restrictive proprietary license. While the terms for API use and Vertex AI integration are clearly documented, the license is not open-source or open-weights. Users are subject to 'Pre-GA Offerings Terms' which allow Google to deprecate or change the model with minimal notice, as seen with the rapid deprecation of the initial 3.0 preview in favor of 3.1.

As a closed API model, there is no official documentation for local VRAM requirements. Third-party reports suggest that running a model of this scale would require at least 80GB of VRAM (e.g., H100/A100) for full context, but Google provides no guidance on quantization tradeoffs or memory scaling for the weights themselves, as they are not available for download.

Google uses clear semantic versioning (3.0 vs 3.1) and provides a public changelog for API updates. The deprecation schedule for preview models is explicitly communicated (e.g., the March 2026 shutdown of the 3.0 preview). However, the 'High' reasoning mode introduces dynamic compute which can lead to variable response behavior, making it harder to track subtle performance drift compared to static models.