Paper · 2023-12-19 · cs.CL, cs.AI, cs.CV
Gemini: A Family of Highly Capable Multimodal Models
Technical report for Gemini 1.0 (Ultra, Pro, Nano): decoder-only Transformers with 32K context and multi-query attention, trained jointly on text, image, audio and video, then post-trained with SFT, reward models and RLHF. Gemini Ultra reports 90.04% on MMLU via uncertainty-routed chain-of-thought (32 samples) and state of the art on 30 of 32 benchmarks.
Multimodal language modelChain-of-thought promptingReinforcement learning from human feedback (RLHF)Multi-query attention (MQA)multimodalllmtechnical-reportpost-trainingevaluationon-device
AI-compiled by claude-opus-5-5 on 2026-10-18, not yet reviewed by a human. Every claim cites its source; check it when a detail matters.
In plain words
Most early AI assistants were built for text and had vision or speech bolted on afterwards, often by turning an image or a recording into a text description first. Gemini 1.0 is Google's family of models trained from the start on text, images, audio and video together, so one model can read a chart, listen to a question and answer in text, or reply with a mix of text and generated images.
The family comes in three sizes. Ultra is the largest and most capable, Pro trades some quality for cost and speed, and Nano is small enough to run on a phone, learning from the bigger models and stored in compressed 4-bit form. After the main training, each model is tuned on human-written example answers and then on human judgments of which answers are better, to make it more helpful, more factual and safer.
The report is mostly a scorecard rather than a recipe. It shows Ultra beating human-expert level on MMLU, a broad exam benchmark, with a trick: sample many step-by-step answers, and only trust the majority vote when the model agrees with itself strongly enough. Almost none of the numbers needed to rebuild the models (layer counts, learning rates, data mixtures) are given, so treat this card as a guide to what was measured and how, not as a training recipe.
TL;DR
Gemini 1.0 is a family of decoder-only Transformers (Ultra, Pro, Nano-1 1.8B, Nano-2 3.25B) trained natively on interleaved text, image, audio and video, with 32K context and multi-query attention §2 tab. 1. The report gives very few reproducible details: no layer counts, widths, optimizer, learning rate or token counts for any size §2 §4. Its reusable contributions are evaluation methodology (uncertainty-routed chain-of-thought on MMLU, decontamination caveats) §5.1.1 §10.2, the post-training pipeline outline (prompts → SFT → reward model → iterative RLHF) §6.3, tool use framed as code generation §6.5.2, and large author-reported benchmark tables tab. 2 tab. 7 tab. 11. Read it to compare against Gemini 1.0 or to copy its evaluation protocols, not to train a model.
Problem
The authors want one model family with strong general ability across modalities together with top reasoning in each domain, covering deployments from datacenter to on-device memory budgets §1 tab. 1. Earlier multimodal systems the paper builds on attach vision to a language model; Gemini is described as multimodal from the beginning and able to emit images as discrete tokens §2. Training at Ultra scale also raises systems problems: hardware failures and silent data corruption become frequent enough that checkpoint-and-restart training loses too much time §3.
Method
Architecture §2
- Backbone: Transformer decoder "enhanced with improvements in architecture and model optimization" for stable training at scale and efficient TPU inference; the improvements are not described §2.
- Context length 32,768 tokens; efficient attention such as multi-query attention §2 §5.1.5.
- Inputs: text interleaved with natural images, charts, screenshots, PDFs, video and audio §2 fig. 2. Visual encoding is "inspired by" Flamingo, CoCa and PaLI, but the model is multimodal from the start of training §2. Encoder architecture, resolution scheme and image tokenizer are not stated.
- Video: encoded as a sequence of frames placed in the context window §2. For the video benchmarks, 16 equally spaced frames per clip §5.2.2.
- Variable input resolution, so more compute can be spent on fine-grained tasks; the mechanism is not stated §2.
- Audio: ingested directly as 16 kHz Universal Speech Model (USM) features, rather than via a text transcript §2.
- Outputs: text, and images via discrete image tokens (citing DALL·E and Parti) §2 §5.2.3. The image decoder is not described.
- Sizes tab. 1:
| Size | Description | Parameters |
|---|---|---|
| Ultra | most capable; serveable at scale on TPUs | not stated |
| Pro | optimized for cost and latency | not stated |
| Nano-1 | on-device, low-memory devices; distilled from larger Gemini models; 4-bit quantized | 1.8B |
| Nano-2 | on-device, high-memory devices; distilled; 4-bit quantized | 3.25B |
- Post-trained variants need "additional architectural modifications" to train multi-objective reward models for RLHF; these are not described §10.1.
Pre-training data §4
- Multimodal, multilingual data: web documents, books, code, plus image, audio and video §4.
- SentencePiece tokenizer trained on a large sample of the whole training corpus; the authors say this helps tokenize non-Latin scripts efficiently §4. Vocabulary size not stated.
- Token budget: largest models follow the Chinchilla approach; smaller models are trained on significantly more tokens for a better inference-cost trade-off, in the spirit of LLaMA §4. Actual token counts not stated.
- Quality filters (heuristic rules and model-based classifiers), safety filtering, and removal of evaluation data found in the training corpus §4.
- Mixture weights chosen by ablations on smaller models; training is staged so domain-relevant data is up-weighted towards the end §4. Weights and stage boundaries not stated.
Training infrastructure §3
- Hardware: TPUv5e and TPUv4, depending on size and configuration; Ultra on a large fleet of TPUv4 across multiple datacenters §3.
- TPUv4 SuperPods of 4096 chips, whose optical switch can rearrange 4x4x4 chip cubes into arbitrary 3D tori in about 10 seconds; for Ultra a few cubes per superpod are held back as hot standbys and for rolling maintenance §3.
- Parallelism: model parallelism within a superpod, data parallelism across superpods, synchronous training over Google's datacenter networks §3.
- Software: JAX and Pathways "single controller" (one Python process drives the whole run); GSPMD partitions the step; the MegaScale XLA pass schedules collectives to overlap with compute §3.
- Fault tolerance: instead of periodic checkpoints to persistent storage, keep redundant in-memory copies of model state and recover from an intact replica after an unplanned failure. Goodput (time on useful new steps over elapsed job time) of the largest job rose from 85% to 97% §3.
- Silent data corruption is expected to hit training every week or two at this scale; handled with deterministic replay to isolate bad computations plus proactive SDC scanners on idle machines and hot standbys §3.
Post-training §6.3 fig. 7
Pipeline: Gemini post-training (Apps and API variants) [§6.3] [§10.1]
Input: pre-trained checkpoint from late in pre-training [§10.1]
1. Prompt collection: curate single- and multi-turn target prompts
(vendor-created, third-party licensed, synthetic). Used for demonstration
and feedback collection, and directly as RL prompts.
2. SFT on demonstrations: target responses written by experts, or generated by
a model and sometimes revised/reviewed by humans; diversity enforced with
data-analysis tools and heuristics across capabilities and semantic clusters.
3. Reward model: human raters give relative preferences over candidate
responses and per-response feedback (creativity, safety, factuality, ...).
Train RMs to match human preferences. Prompt selection and the sampling
strategy for candidates strongly affect the data's usefulness.
4. RLHF against the RM. Iterate: RL exploits the RM, the RM is improved via
evaluation and new data collection, repeat.
Output: Gemini Apps (chat) or Gemini API (developer) modelRL algorithm, KL penalty, learning rates and data sizes are not stated.
Capability-specific additions §6.5:
- Instruction following: data across instruction categories; for programmatically verifiable constraints (e.g. word count) synthetic data is made by prompting and editing responses so the constraint holds §6.5.1.
- Tool use as code generation: each tool invocation is a code block that calls tools; a block may compose several tools, and the model sees execution results before continuing §6.5.2 fig. 8:
Loop: tool use at inference [§6.5.2] [fig. 8]
1. Sample from the LLM given the user prompt and context so far.
2. If the sample contains a tool code block: execute it, append the results
to the context, go to 1.
3. Otherwise return the response.
(Stopping rule and limits on iterations: not stated.)- Multilingual (API): add human-written non-English pairs and machine-translated English demonstrations; drop pairs that do not survive translation (strict word counts, English-specific forms like iambic pentameter, Latin-script letter puzzles); each translated pair is rated by at least 3 raters and kept if the majority rate it accurate §6.5.3.
- Multilingual (Apps): localize English data to native cultures §6.5.3.
- Vision: fine-tune on a balanced mixture of text-only and image-text data §6.5.4; safety SFT for text-only queries did not transfer to image+text harm-inducing queries, so separate multimodal safety SFT sets were built §7.3.2.
- Safety SFT: about 20 harm types; harm-inducing queries come from experts, LLM generation from policy text plus seed keywords, and automated red teaming; responses are produced by a Constitutional-AI-like recipe where variants of Google's content policy act as constitutions for model revision and candidate selection §7.3.2. Safety preference data is mixed into RM training §7.3.2.
Uncertainty-routed chain-of-thought (MMLU) §5.1.1 §10.2
Algorithm: uncertainty-routed CoT@k [§5.1.1] [§10.2] [tab. 2 note *]
Input: question q, model M, sample count k (8 or 32), threshold t
1. Draw k chain-of-thought samples from M for q; extract each final answer.
2. Let a* be the majority answer and c its share of votes (consensus).
3. If c > t: return a*.
4. Else: return the greedy, maximum-likelihood answer without chain of thought.
t is tuned per model on the validation split.The exact consensus measure, the value of t, sampling temperature and the prompt are not stated. Rationale: when the model is visibly inconsistent, CoT samples can do worse than the maximum-likelihood choice §10.2.
Hyperparameters
| Name | Value | Where | Notes |
|---|---|---|---|
| context length | 32,768 tokens | §2 §5.1.5 | all sizes |
| attention | efficient, e.g. multi-query | §2 | whether every size uses MQA not stated |
| Nano-1 / Nano-2 params | 1.8B / 3.25B | tab. 1 §5.1.3 | |
| Ultra / Pro params | not stated | §10.1 | model card: "Model Stats: Not reported" |
| layers, width, heads, vocab | not stated | ||
| tokenizer | SentencePiece, trained on a large corpus sample | §4 | |
| training tokens | not stated | §4 | largest models Chinchilla-style; small ones over-trained |
| optimizer, LR, schedule, batch | not stated | ||
| initialization | random (pre-training); post-training from late pre-training checkpoints | §10.1 | |
| Nano deployment precision | 4-bit | tab. 1 | quantization method not stated |
| audio input | USM features, 16 kHz | §2 | |
| video frames (eval) | 16 equally spaced per clip | §5.2.2 | |
| CoT samples k | 32 (Ultra), 8 (Pro) on MMLU | tab. 2 | |
| CoT routing threshold | tuned on validation split; value not stated | §10.2 | |
| GSM8K decoding | CoT + self-consistency, Maj1@32 | §5.1.1 tab. 2 | |
| hardware | TPUv4, TPUv5e | §3 §10.1 | compute "Not reported" §10.1 |
| translated-data filter | ≥3 raters, majority must rate accurate | §6.5.3 | |
| safety harm types | ~20 | §7.3.2 | |
| compute requirements | not reported | §10.1 |
Results
All numbers are author-reported. Unless marked "PT" (post-trained Gemini API model), text results are for pre-trained models; GPT-4 / GPT-3.5 entries marked "API" were self-collected via the API in November 2023, others are as reported by their authors tab. 2.
Text benchmarks tab. 2
| Benchmark | Ultra | Pro | GPT-4 | GPT-3.5 | PaLM 2-L | Claude 2 | Inflection-2 | Grok 1 | LLAMA-2 |
|---|---|---|---|---|---|---|---|---|---|
| MMLU | 90.04% CoT@32*; 83.7% 5-shot | 79.13% CoT@8*; 71.8% 5-shot | 87.29% CoT@32 (API); 86.4% 5-shot (reported) | 70% 5-shot | 78.4% 5-shot | 78.5% 5-shot CoT | 79.6% 5-shot | 73.0% 5-shot | 68.0%*** |
| GSM8K | 94.4% Maj1@32 | 86.5% Maj1@32 | 92.0% SFT & 5-shot CoT | 57.1% 5-shot | 80.0% 5-shot | 88.0% 0-shot | 81.4% 8-shot | 62.9% 8-shot | 56.8% 5-shot |
| MATH | 53.2% 4-shot | 32.6% 4-shot | 52.9% 4-shot (API); 50.3% (Zheng et al.) | 34.1% 4-shot (API) | 34.4% 4-shot | — | 34.8% | 23.9% 4-shot | 13.5% 4-shot |
| BIG-Bench-Hard | 83.6% 3-shot | 75.0% 3-shot | 83.1% 3-shot (API) | 66.6% 3-shot (API) | 77.7% 3-shot | — | — | — | 51.2% 3-shot |
| HumanEval | 74.4% 0-shot (PT) | 67.7% 0-shot (PT) | 67.0% 0-shot (reported) | 48.1% 0-shot | — | 70.0% 0-shot | 44.5% 0-shot | 63.2% 0-shot | 29.9% 0-shot |
| Natural2Code | 74.9% 0-shot | 69.6% 0-shot | 73.9% 0-shot (API) | 62.3% 0-shot (API) | — | — | — | — | — |
| DROP (F1) | 82.4 variable shots | 74.1 variable shots | 80.9 3-shot (reported) | 64.1 3-shot | 82.0 variable shots | — | — | — | — |
| HellaSwag (val) | 87.8% 10-shot | 84.7% 10-shot | 95.3% 10-shot (reported) | 85.5% 10-shot | 86.8% 10-shot | — | 89.0% 10-shot | — | 80.0%*** |
| WMT23 (BLEURT) | 74.4 1-shot (PT) | 71.7 1-shot | 73.8 1-shot (API) | — | 72.7 1-shot | — | — | — | — |
* uncertainty-routed CoT (see Method). *** decontaminated numbers from the LLAMA-2 report tab. 2.
Other text results:
- MMLU routing ablation §10.2 fig. 9: Gemini Ultra 84.0% greedy → 85.0% with 32 CoT samples only → 90.0% uncertainty-routed CoT@32. GPT-4 84.2% greedy → 87.3% uncertainty-routed CoT@32, a gain it already gets from 32 CoT samples alone. Human-expert MMLU level is 89.8% per the benchmark authors; prior SOTA 86.4% §5.1.1.
- AMC 2022–2023 (150 questions): Ultra 32% vs GPT-4 30%; smaller models near random §5.1.1.
- AlphaCode 2 (built on a fine-tuned Gemini Pro used as both generator and reward model, with filtering, clustering and reranking): 43% of 77 problems from 12 Codeforces contests vs 25% for AlphaCode; estimated 85th percentile §5.1.7.
- Long context: synthetic key-value retrieval across the full context, Ultra 98% accuracy; NLL on held-out long documents decreases with position up to 32K §5.1.5 fig. 4.
Nano tab. 3
| Task | Nano-1 acc. | Nano-1 / Pro | Nano-2 acc. | Nano-2 / Pro |
|---|---|---|---|---|
| BoolQ | 71.6 | 0.81 | 79.3 | 0.90 |
| TydiQA (GoldP) | 68.9 | 0.85 | 74.2 | 0.91 |
| NaturalQuestions (Retrieved) | 38.6 | 0.69 | 46.5 | 0.83 |
| NaturalQuestions (Closed-book) | 18.8 | 0.43 | 24.8 | 0.56 |
| BIG-Bench-Hard (3-shot) | 34.8 | 0.47 | 42.4 | 0.58 |
| MBPP | 20.0 | 0.33 | 27.2 | 0.45 |
| MATH (4-shot) | 13.5 | 0.41 | 22.8 | 0.70 |
| MMLU (5-shot) | 45.9 | 0.64 | 55.8 | 0.78 |
Multilingual tab. 4 tab. 5 §5.1.4
| WMT 23, avg BLEURT, 1-shot | Ultra | Pro | Nano-2 | Nano-1 | GPT-4 | PaLM 2-L |
|---|---|---|---|---|---|---|
| High resource | 74.2 | 71.7 | 67.7 | 64.1 | 74.0 | 72.6 |
| Mid resource | 74.7 | 71.8 | 67.0 | 64.8 | 73.6 | 72.7 |
| Out-of-English | 74.8 | 71.5 | 66.2 | 65.2 | 73.6 | 72.2 |
| Into-English | 73.9 | 72.0 | 69.0 | 63.5 | 74.1 | 73.4 |
| All languages | 74.4 | 71.7 | 67.4 | 64.8 | 73.8 | 72.7 |
The WMT evaluation used a post-trained Gemini API Ultra model §5.1.4.
| Task | Ultra | Pro | GPT-4 | PaLM 2-L |
|---|---|---|---|---|
| MGSM (8-shot) | 79.0 | 63.5 | 74.5 | 74.7 |
| XLSum (3-shot, ROUGE-L) | 17.6 | 16.2 | — | 15.4 |
| WikiLingua (BLEURT) | 48.9 | 47.8 | — | 50.4 |
Very-low-resource languages (Tamazight, Kanure, North Ndebele, Quechua), to and from English, 1-shot chrF: Ultra 27.0 vs PaLM 2-L 25.3 §5.1.4. WikiLingua compares Ultra 5-shot with PaLM 2 3-shot §5.1.4.
Factuality, Gemini API Pro tab. 6
| Variant | Inaccurate rate (↓) | Attribution AIS (↑) | Hedging acc. (↑) |
|---|---|---|---|
| No factuality-focused adaptation | 6.7% [5.8, 7.8] | 40.2% [37.9, 42.5] | 0% |
| Final post-training stage | 3.8% [3.1, 4.8] | 60.0% [57.6, 62.1] | 69.3% |
Brackets are 95% confidence intervals. Factuality and attribution are judged by human annotators; hedging is evaluated automatically §5.1.6.
Image understanding, post-trained Gemini API, 0-shot, pixel only tab. 7
| Benchmark | Ultra | Pro | Nano-2 | Nano-1 | GPT-4V | Prior SOTA |
|---|---|---|---|---|---|---|
| MMMU (val) | 59.4% pass@1; 62.4% Maj1@32 | 47.9% | 32.6% | 26.3% | 56.8% | 56.8% (GPT-4V, 0-shot) |
| TextVQA (val) | 82.3% | 74.6% | 65.9% | 62.5% | 78.0% | 79.5% (PaLI-3, fine-tuned) |
| DocVQA (test) | 90.9% | 88.1% | 74.3% | 72.2% | 88.4% (pixel only) | 88.4% (GPT-4V, 0-shot) |
| ChartQA (test) | 80.8% | 74.1% | 51.9% | 53.6% | 78.5% (4-shot CoT) | 79.3% (DePlot, 1-shot PoT) |
| InfographicVQA (test) | 80.3% | 75.2% | 54.5% | 51.1% | 75.1% (pixel only) | 75.1% (GPT-4V, 0-shot) |
| MathVista (testmini) | 53.0% | 45.2% | 30.6% | 27.3% | 49.9% | 49.9% (GPT-4V, 0-shot) |
| AI2D (test) | 79.5% | 73.9% | 51.0% | 37.9% | 78.2% | 81.4% (PaLI-X, fine-tuned) |
| VQAv2 (test-dev) | 77.8% | 71.2% | 67.5% | 62.7% | 77.2% | 86.1% (PaLI-X, fine-tuned) |
Greedy decoding, no external OCR §5.2.1. MathVista used the benchmark authors' evaluation script §5.2.1.
MMMU per discipline tab. 8:
| Discipline | Ultra Maj@32 | Ultra pass@1 | GPT-4V pass@1 |
|---|---|---|---|
| Art & Design | 74.2 | 70.0 | 65.8 |
| Business | 62.7 | 56.7 | 59.3 |
| Science | 49.3 | 48.0 | 54.7 |
| Health & Medicine | 71.3 | 67.3 | 64.7 |
| Humanities & Social Science | 78.3 | 78.3 | 72.5 |
| Technology & Engineering | 53.0 | 47.1 | 36.7 |
| Overall | 62.4 | 59.4 | 56.8 |
XM-3600 captioning, 4-shot, CIDEr, Flamingo protocol tab. 9 §5.2.1:
| Language | Ultra | Pro | PaLI-X |
|---|---|---|---|
| English | 86.4 | 87.1 | 77.8 |
| French | 77.9 | 76.7 | 62.5 |
| Hindi | 31.1 | 29.8 | 22.2 |
| Modern Hebrew | 54.5 | 52.6 | 38.7 |
| Romanian | 39.0 | 37.7 | 30.2 |
| Thai | 86.7 | 77.0 | 56.0 |
| Chinese | 33.3 | 30.2 | 27.7 |
| Average (of 7) | 58.4 | 55.9 | 45.0 |
Pre-trained vs post-trained Ultra, 0-shot tab. 17: TextVQA 81.4 → 82.3, DocVQA 90.1 → 90.9, ChartQA 80.8 → 80.8, InfographicVQA 77.9 → 80.3, AI2D 76.6 → 79.5, VQAv2 74.5 → 77.8 (MMMU and MathVista: pre-trained n/a).
Video tab. 10
| Task (metric) | Ultra | Pro | Few-shot SoTA |
|---|---|---|---|
| VATEX test (CIDEr, 4-shot) | 62.7 | 57.4 | 56.0 (Flamingo, 4-shot) |
| VATEX ZH test (CIDEr, 4-shot) | 51.3 | 50.0 | — |
| YouCook2 val (CIDEr, 4-shot) | 135.4 | 123.2 | 74.5 (Flamingo, 4-shot) |
| NextQA test (WUPS, 0-shot) | 29.9 | 28.0 | 26.7 (Flamingo, 0-shot) |
| ActivityNet-QA test (top-1, 0-shot) | 52.2 | 49.8 | 45.3 (Video-LLAVA, 0-shot) |
| Perception Test MCQA test (top-1, 0-shot) | 54.7 | 51.1 | 46.3 (SeViLA, 0-shot) |
Audio tab. 11
| Task | Metric | Pro | Nano-1 | Whisper | USM |
|---|---|---|---|---|---|
| YouTube (en-us) | WER ↓ | 4.9% | 5.5% | 6.5% (v3) | 6.2% |
| Multilingual Librispeech (en-us) | WER ↓ | 4.8% | 5.9% | 6.2% (v2) | 7.0% |
| FLEURS (62 lang) | WER ↓ | 7.6% | 14.2% | 17.6% (v3) | 11.8% |
| VoxPopuli (14 lang) | WER ↓ | 9.1% | 9.5% | 15.9% (v2) | 13.4% |
| CoVoST 2 (21 lang) | BLEU ↑ | 40.1 | 35.4 | 29.1 (v2) | 30.7 |
Ultra was not evaluated on audio §5.2.4.
Post-training §6.5
| Evaluation | Result | Source |
|---|---|---|
| Complex-prompt IF, per-instruction acc. | PaLM 2 (post-trained) 59.5±3.0%, Gemini (Pro) 77.8±2.0%, Gemini Advanced (Ultra) 87.4±1.4% | tab. 14 |
| Complex-prompt IF, full-response acc. | 25.5±3.3%, 38.5±3.6%, 54.1±3.7% (same order) | tab. 14 |
| Tools (early Gemini API Pro), with vs without | GSM8K 80.1 vs 69.7, MATH 41.8 vs 30.7, NQ 68.0 vs 59.0, Realtime QA 70.8 vs 39.2 | tab. 15 |
| Gemini Apps with tools vs without (internal, excl. ties) | preferred 78% of the time | §6.5.2 |
| Multilingual Gemini (Pro) vs older PaLM 2 recipe: SxS / MBPP / MMLU | ja-JP +0.14 / +22.2% / +3.6%; pt-BR +0.17 / +23.2% / +5.2%; de-DE +0.1 / +21.4% / +7.5%; es-419 +0.12 / +22.8% / +9.3%; it-IT +0.13 / +13.8% / +7.5% | tab. 16 |
| Image-text data, text-only SxS (Apps Pro) | +0.01±0.01 vs text-only training | §6.5.4 |
| SFT+RLHF vs SFT, image-understanding SxS (Apps Pro) | +0.223±0.06 | §6.5.4 |
| Internal coding SxS | Gemini (Pro) vs Bard (PaLM 2, Sept. 2023) 0.19±0.03; Gemini Advanced (Ultra) vs Gemini (Pro) 0.13±0.02 | tab. 18 |
SxS score: each side-by-side rating is mapped to an ordinal value in [−1.5, 1.5], positive when model A is preferred, then averaged §6.5.3.
Safety §7.4
- BBQ average bias score stays near 0 on a −1 (stereotype-countering) to +1 (stereotype-reinforcing) scale; average Real Toxicity Prompts toxicity during pre-training fluctuates around 6% (10k non-toxic prompts, Perspective API) §7.4.1.
- Winogender, Winobias and BBQ mostly saturate near 99% accuracy §7.4.1.
- Dangerous-capability evaluations (cyber CTFs, persuasion, self-proliferation, situational awareness, CBRN) are summarized qualitatively, without numbers §7.4.1.
Paper vs code
No official code.
Pitfalls
- Not a reproducible recipe. Parameter counts for Pro and Ultra, depth, width, heads, vocabulary, optimizer, learning rate, batch size, token counts, data mixture weights, the visual encoder, the image decoder and the RL algorithm are all unstated §2 §4 §6.3; the model card lists compute and model stats as "Not reported" §10.1. Do not fill these from later Gemini reports or general knowledge.
- The 90.04% MMLU headline is not 5-shot. It needs 32 CoT samples plus a routing threshold tuned on validation data; plain 5-shot Ultra is 83.7%, below GPT-4's reported 86.4% 5-shot tab. 2. Pro's 79.13% uses k=8, not 32 tab. 2. The threshold value and consensus definition are never given §10.2, so the routing cannot be reproduced exactly.
- Ambiguous GPT-4 MMLU entry. The table shows GPT-4 87.29% "CoT@32 (via API)" tab. 2, while the appendix gives 87.3% for GPT-4 with uncertainty-routed CoT@32 and says plain 32-sample CoT already reaches that §10.2. Which method the table entry used is unclear.
- Mixed evaluation settings in tab. 2. Columns mix pre-trained and post-trained (PT) models, self-collected API runs and numbers taken from other reports, different shot counts, and "variable shots" for DROP tab. 2. HumanEval and WMT23 Ultra numbers are post-trained; most others are pre-trained. Compare row by row, not as a uniform leaderboard.
- Contamination. About a hundred extra fine-tuning steps on website extracts matching the HellaSwag training set raise 1-shot validation accuracy to 89.6% (Pro) and 96.0% (Ultra), so the authors report only 10-shot decontaminated HellaSwag §5.1.1. LAMBADA results were withheld because of leakage §5.1.1, yet LAMBADA still appears in the Reasoning harness behind the capability plot §10.3 fig. 3. The decontamination method is not described §4.
- FLEURS trained on its own training set. Pro's 7.6% FLEURS WER benefits from FLEURS training data; without it the WER is 15.8 (still better than Whisper's 17.6) §5.2.4 tab. 11.
- Video evaluation subset. YouTube-based video datasets were scored only on clips still public in November 2023, which may differ from the subsets other models used §5.2.2. Only 16 frames per clip are sampled §5.2.2.
- Image comparisons are not apples-to-apples. Several "prior SOTA" entries are fine-tuned models, while Gemini is 0-shot; Gemini still trails fine-tuned PaLI-X on VQAv2 (77.8 vs 86.1) and AI2D (79.5 vs 81.4) tab. 7. On MMMU, Ultra is below GPT-4V in Science (49.3 vs 54.7) tab. 8. The 62.4% MMMU headline is Maj1@32; pass@1 is 59.4% tab. 7.
- Small numeric inconsistencies in the text. Hedging is "70%" in the text but 69.3% in the table §5.1.6 tab. 6; MMLU appears as 90.04% §5.1.1 and 90.0% §8.
- Nano details missing. Nano is distilled from larger Gemini models and 4-bit quantized, but neither the distillation objective nor the quantization scheme is described tab. 1 §2. Its summarization and reading-comprehension strength relies on per-task fine-tuning §5.1.3.
- Infrastructure costs hidden by the abstract. Ultra spans multiple datacenters; silent data corruption is expected every one to two weeks; the 85% → 97% goodput gain depends on keeping redundant in-memory replicas of the model state, which costs memory and hardware headroom §3.
- Tool use is code. Tool calls are emitted as code blocks the system runs, so a reimplementation needs a sandboxed executor and a loop that feeds results back; the stopping rule and safety constraints are not stated §6.5.2 fig. 8.
- Safety evaluation results are mostly qualitative. Content-safety, red-teaming and dangerous-capability findings are given without violation rates §7.4.
Lineage
- Builds on: Transformer decoders; multi-query attention (arXiv:1911.02150) §2; Flamingo (arXiv:2204.14198), CoCa (arXiv:2205.01917) and PaLI (arXiv:2209.06794) for visual encoding §2; USM (arXiv:2303.01037) audio features §2; Chinchilla (arXiv:2203.15556) and LLaMA (arXiv:2302.13971) token budgets §4; chain-of-thought (arXiv:2201.11903) and self-consistency (arXiv:2203.11171) prompting §5.1.1; PaLM and PaLM 2 (arXiv:2305.10403) as Google's previous flagship LLMs and infrastructure baseline §3.
- Compared against: GPT-4, GPT-4V, GPT-3.5, PaLM 2-L, Claude 2, Inflection-2, Grok 1, LLAMA-2, Whisper, USM, PaLI-X, Flamingo tab. 2 tab. 7 tab. 10 tab. 11.
- Superseded by: not known from content in this wiki.
Source and license
Adapted from “Gemini: A Family of Highly Capable Multimodal Models” by Gemini Team, Rohan Anil, Sebastian Borgeaud, Jean-Baptiste Alayrac, Jiahui Yu, Radu Soricut, Johan Schalkwyk, Andrew M. Dai, Anja Hauth, Katie Millican, David Silver, Melvin Johnson, Ioannis Antonoglou, Julian Schrittwieser, Amelia Glaese, Jilin Chen, Emily Pitler, Timothy Lillicrap, Angeliki Lazaridou, Orhan Firat, James Molloy, Michael Isard, Paul R. Barham, Tom Hennigan, Benjamin Lee, Fabio Viola, Malcolm Reynolds, Yuanzhong Xu, Ryan Doherty, Eli Collins, Clemens Meyer, Eliza Rutherford, Erica Moreira, Kareem Ayoub, Megha Goel, Jack Krawczyk, Cosmo Du, Ed Chi, Heng-Tze Cheng, Eric Ni, Purvi Shah, Patrick Kane, Betty Chan, Manaal Faruqui, Aliaksei Severyn, Hanzhao Lin, YaGuang Li, Yong Cheng, Abe Ittycheriah, Mahdis Mahdieh, Mia Chen, Pei Sun, Dustin Tran, Sumit Bagri, Balaji Lakshminarayanan, Jeremiah Liu, Andras Orban, Fabian Güra, Hao Zhou, Xinying Song, Aurelien Boffy, Harish Ganapathy, Steven Zheng, HyunJeong Choe, Ágoston Weisz, Tao Zhu, Yifeng Lu, Siddharth Gopal, Jarrod Kahn, Maciej Kula, Jeff Pitman, Rushin Shah, Emanuel Taropa, Majd Al Merey, Martin Baeuml, Zhifeng Chen, Laurent El Shafey, Yujing Zhang, Olcan Sercinoglu, George Tucker, Enrique Piqueras, Maxim Krikun, Iain Barr, Nikolay Savinov, Ivo Danihelka, Becca Roelofs, Anaïs White, Anders Andreassen, Tamara von Glehn, Lakshman Yagati, Mehran Kazemi, Lucas Gonzalez, Misha Khalman, Jakub Sygnowski, Alexandre Frechette, Charlotte Smith, Laura Culp, Lev Proleev, Yi Luan, Xi Chen, James Lottes, Nathan Schucher, Federico Lebron, Alban Rrustemi, Natalie Clay, Phil Crone, Tomas Kocisky, Jeffrey Zhao, Bartek Perz, Dian Yu, Heidi Howard, Adam Bloniarz, Jack W. Rae, Han Lu, Laurent Sifre, Marcello Maggioni, Fred Alcober, Dan Garrette, Megan Barnes, Shantanu Thakoor, Jacob Austin, Gabriel Barth-Maron, William Wong, Rishabh Joshi, Rahma Chaabouni, Deeni Fatiha, Arun Ahuja, Gaurav Singh Tomar, Evan Senter, Martin Chadwick, Ilya Kornakov, Nithya Attaluri, Iñaki Iturrate, Ruibo Liu, Yunxuan Li, Sarah Cogan, Jeremy Chen, Chao Jia, Chenjie Gu, Qiao Zhang, Jordan Grimstad, Ale Jakse Hartman, Xavier Garcia, Thanumalayan Sankaranarayana Pillai, Jacob Devlin, Michael Laskin, Diego de Las Casas, Dasha Valter, Connie Tao, Lorenzo Blanco, Adrià Puigdomènech Badia, David Reitter, Mianna Chen, Jenny Brennan, Clara Rivera, Sergey Brin, Shariq Iqbal, Gabriela Surita, Jane Labanowski, Abhi Rao, Stephanie Winkler, Emilio Parisotto, Yiming Gu, Kate Olszewska, Ravi Addanki, Antoine Miech, Annie Louis, Denis Teplyashin, Geoff Brown, Elliot Catt, Jan Balaguer, Jackie Xiang, Pidong Wang, Zoe Ashwood, Anton Briukhov, Albert Webson, Sanjay Ganapathy, Smit Sanghavi, Ajay Kannan, Ming-Wei Chang, Axel Stjerngren, Josip Djolonga, Yuting Sun, Ankur Bapna, Matthew Aitchison, Pedram Pejman, Henryk Michalewski, Tianhe Yu, Cindy Wang, Juliette Love, Junwhan Ahn, Dawn Bloxwich, Kehang Han, Peter Humphreys, Thibault Sellam, James Bradbury, Varun Godbole, Sina Samangooei, Bogdan Damoc, Alex Kaskasoli, Sébastien M. R. Arnold, Vijay Vasudevan, Shubham Agrawal, Jason Riesa, Dmitry Lepikhin, Richard Tanburn, Srivatsan Srinivasan, Hyeontaek Lim, Sarah Hodkinson, Pranav Shyam, Johan Ferret, Steven Hand, Ankush Garg, Tom Le Paine, Jian Li, Yujia Li, Minh Giang, Alexander Neitz, Zaheer Abbas, Sarah York, Machel Reid, Elizabeth Cole, Aakanksha Chowdhery, Dipanjan Das, Dominika Rogozińska, Vitaliy Nikolaev, Pablo Sprechmann, Zachary Nado, Lukas Zilka, Flavien Prost, Luheng He, Marianne Monteiro, Gaurav Mishra, Chris Welty, Josh Newlan, Dawei Jia, Miltiadis Allamanis, Clara Huiyi Hu, Raoul de Liedekerke, Justin Gilmer, Carl Saroufim, Shruti Rijhwani, Shaobo Hou, Disha Shrivastava, Anirudh Baddepudi, Alex Goldin, Adnan Ozturel, Albin Cassirer, Yunhan Xu, Daniel Sohn, Devendra Sachan, Reinald Kim Amplayo, Craig Swanson, Dessie Petrova, Shashi Narayan, Arthur Guez, Siddhartha Brahma, Jessica Landon, Miteyan Patel, Ruizhe Zhao, Kevin Villela, Luyu Wang, Wenhao Jia, Matthew Rahtz, Mai Giménez, Legg Yeung, James Keeling, Petko Georgiev, Diana Mincu, Boxi Wu, Salem Haykal, Rachel Saputro, Kiran Vodrahalli, James Qin, Zeynep Cankara, Abhanshu Sharma, Nick Fernando, Will Hawkins, Behnam Neyshabur, Solomon Kim, Adrian Hutter, Priyanka Agrawal, Alex Castro-Ros, George van den Driessche, Tao Wang, Fan Yang, Shuo-yiin Chang, Paul Komarek, Ross McIlroy, Mario Lučić, Guodong Zhang, Wael Farhan, Michael Sharman, Paul Natsev, Paul Michel, Yamini Bansal, Siyuan Qiao, Kris Cao, Siamak Shakeri, Christina Butterfield, Justin Chung, Paul Kishan Rubenstein, Shivani Agrawal, Arthur Mensch, Kedar Soparkar, Karel Lenc, Timothy Chung, Aedan Pope, Loren Maggiore, Jackie Kay, Priya Jhakra, Shibo Wang, Joshua Maynez, Mary Phuong, Taylor Tobin, Andrea Tacchetti, Maja Trebacz, Kevin Robinson, Yash Katariya, Sebastian Riedel, Paige Bailey, Kefan Xiao, Nimesh Ghelani, Lora Aroyo, Ambrose Slone, Neil Houlsby, Xuehan Xiong, Zhen Yang, Elena Gribovskaya, Jonas Adler, Mateo Wirth, Lisa Lee, Music Li, Thais Kagohara, Jay Pavagadhi, Sophie Bridgers, Anna Bortsova, Sanjay Ghemawat, Zafarali Ahmed, Tianqi Liu, Richard Powell, Vijay Bolina, Mariko Iinuma, Polina Zablotskaia, James Besley, Da-Woon Chung, Timothy Dozat, Ramona Comanescu, Xiance Si, Jeremy Greer, Guolong Su, Martin Polacek, Raphaël Lopez Kaufman, Simon Tokumine, Hexiang Hu, Elena Buchatskaya, Yingjie Miao, Mohamed Elhawaty, Aditya Siddhant, Nenad Tomasev, Jinwei Xing, Christina Greer, Helen Miller, Shereen Ashraf, Aurko Roy, Zizhao Zhang, Ada Ma, Angelos Filos, Milos Besta, Rory Blevins, Ted Klimenko, Chih-Kuan Yeh, Soravit Changpinyo, Jiaqi Mu, Oscar Chang, Mantas Pajarskas, Carrie Muir, Vered Cohen, Charline Le Lan, Krishna Haridasan, Amit Marathe, Steven Hansen, Sholto Douglas, Rajkumar Samuel, Mingqiu Wang, Sophia Austin, Chang Lan, Jiepu Jiang, Justin Chiu, Jaime Alonso Lorenzo, Lars Lowe Sjösund, Sébastien Cevey, Zach Gleicher, Thi Avrahami, Anudhyan Boral, Hansa Srinivasan, Vittorio Selo, Rhys May, Konstantinos Aisopos, Léonard Hussenot, Livio Baldini Soares, Kate Baumli, Michael B. Chang, Adrià Recasens, Ben Caine, Alexander Pritzel, Filip Pavetic, Fabio Pardo, Anita Gergely, Justin Frye, Vinay Ramasesh, Dan Horgan, Kartikeya Badola, Nora Kassner, Subhrajit Roy, Ethan Dyer, Víctor Campos Campos, Alex Tomala, Yunhao Tang, Dalia El Badawy, Elspeth White, Basil Mustafa, Oran Lang, Abhishek Jindal, Sharad Vikram, Zhitao Gong, Sergi Caelles, Ross Hemsley, Gregory Thornton, Fangxiaoyu Feng, Wojciech Stokowiec, Ce Zheng, Phoebe Thacker, Çağlar Ünlü, Zhishuai Zhang, Mohammad Saleh, James Svensson, Max Bileschi, Piyush Patil, Ankesh Anand, Roman Ring, Katerina Tsihlas, Arpi Vezer, Marco Selvi, Toby Shevlane, Mikel Rodriguez, Tom Kwiatkowski, Samira Daruki, Keran Rong, Allan Dafoe, Nicholas FitzGerald, Keren Gu-Lemberg, Mina Khan, Lisa Anne Hendricks, Marie Pellat, Vladimir Feinberg, James Cobon-Kerr, Tara Sainath, Maribeth Rauh, Sayed Hadi Hashemi, Richard Ives, Yana Hasson, Eric Noland, Yuan Cao, Nathan Byrd, Le Hou, Qingze Wang, Thibault Sottiaux, Michela Paganini, Jean-Baptiste Lespiau, Alexandre Moufarek, Samer Hassan, Kaushik Shivakumar, Joost van Amersfoort, Amol Mandhane, Pratik Joshi, Anirudh Goyal, Matthew Tung, Andrew Brock, Hannah Sheahan, Vedant Misra, Cheng Li, Nemanja Rakićević, Mostafa Dehghani, Fangyu Liu, Sid Mittal, Junhyuk Oh, Seb Noury, Eren Sezener, Fantine Huot, Matthew Lamm, Nicola De Cao, Charlie Chen, Sidharth Mudgal, Romina Stella, Kevin Brooks, Gautam Vasudevan, Chenxi Liu, Mainak Chain, Nivedita Melinkeri, Aaron Cohen, Venus Wang, Kristie Seymore, Sergey Zubkov, Rahul Goel, Summer Yue, Sai Krishnakumaran, Brian Albert, Nate Hurley, Motoki Sano, Anhad Mohananey, Jonah Joughin, Egor Filonov, Tomasz Kępa, Yomna Eldawy, Jiawern Lim, Rahul Rishi, Shirin Badiezadegan, Taylor Bos, Jerry Chang, Sanil Jain, Sri Gayatri Sundara Padmanabhan, Subha Puttagunta, Kalpesh Krishna, Leslie Baker, Norbert Kalb, Vamsi Bedapudi, Adam Kurzrok, Shuntong Lei, Anthony Yu, Oren Litvin, Xiang Zhou, Zhichun Wu, Sam Sobell, Andrea Siciliano, Alan Papir, Robby Neale, Jonas Bragagnolo, Tej Toor, Tina Chen, Valentin Anklin, Feiran Wang, Richie Feng, Milad Gholami, Kevin Ling, Lijuan Liu, Jules Walter, Hamid Moghaddam, Arun Kishore, Jakub Adamek, Tyler Mercado, Jonathan Mallinson, Siddhinita Wandekar, Stephen Cagle, Eran Ofek, Guillermo Garrido, Clemens Lombriser, Maksim Mukha, Botu Sun, Hafeezul Rahman Mohammad, Josip Matak, Yadi Qian, Vikas Peswani, Pawel Janus, Quan Yuan, Leif Schelin, Oana David, Ankur Garg, Yifan He, Oleksii Duzhyi, Anton Älgmyr, Timothée Lottaz, Qi Li, Vikas Yadav, Luyao Xu, Alex Chinien, Rakesh Shivanna, Aleksandr Chuklin, Josie Li, Carrie Spadine, Travis Wolfe, Kareem Mohamed, Subhabrata Das, Zihang Dai, Kyle He, Daniel von Dincklage, Shyam Upadhyay, Akanksha Maurya, Luyan Chi, Sebastian Krause, Khalid Salama, Pam G Rabinovitch, Pavan Kumar Reddy M, Aarush Selvan, Mikhail Dektiarev, Golnaz Ghiasi, Erdem Guven, Himanshu Gupta, Boyi Liu, Deepak Sharma, Idan Heimlich Shtacher, Shachi Paul, Oscar Akerlund, François-Xavier Aubet, Terry Huang, Chen Zhu, Eric Zhu, Elico Teixeira, Matthew Fritze, Francesco Bertolini, Liana-Eleonora Marinescu, Martin Bölle, Dominik Paulus, Khyatti Gupta, Tejasi Latkar, Max Chang, Jason Sanders, Roopa Wilson, Xuewei Wu, Yi-Xuan Tan, Lam Nguyen Thiet, Tulsee Doshi, Sid Lall, Swaroop Mishra, Wanming Chen, Thang Luong, Seth Benjamin, Jasmine Lee, Ewa Andrejczuk, Dominik Rabiej, Vipul Ranjan, Krzysztof Styrc, Pengcheng Yin, Jon Simon, Malcolm Rose Harriott, Mudit Bansal, Alexei Robsky, Geoff Bacon, David Greene, Daniil Mirylenka, Chen Zhou, Obaid Sarvana, Abhimanyu Goyal, Samuel Andermatt, Patrick Siegler, Ben Horn, Assaf Israel, Francesco Pongetti, Chih-Wei "Louis" Chen, Marco Selvatici, Pedro Silva, Kathie Wang, Jackson Tolins, Kelvin Guu, Roey Yogev, Xiaochen Cai, Alessandro Agostini, Maulik Shah, Hung Nguyen, Noah Ó Donnaile, Sébastien Pereira, Linda Friso, Adam Stambler, Adam Kurzrok, Chenkai Kuang, Yan Romanikhin, Mark Geller, ZJ Yan, Kane Jang, Cheng-Chun Lee, Wojciech Fica, Eric Malmi, Qijun Tan, Dan Banica, Daniel Balle, Ryan Pham, Yanping Huang, Diana Avram, Hongzhi Shi, Jasjot Singh, Chris Hidey, Niharika Ahuja, Pranab Saxena, Dan Dooley, Srividya Pranavi Potharaju, Eileen O'Neill, Anand Gokulchandran, Ryan Foley, Kai Zhao, Mike Dusenberry, Yuan Liu, Pulkit Mehta, Ragha Kotikalapudi, Chalence Safranek-Shrader, Andrew Goodman, Joshua Kessinger, Eran Globen, Prateek Kolhar, Chris Gorgolewski, Ali Ibrahim, Yang Song, Ali Eichenbaum, Thomas Brovelli, Sahitya Potluri, Preethi Lahoti, Cip Baetu, Ali Ghorbani, Charles Chen, Andy Crawford, Shalini Pal, Mukund Sridhar, Petru Gurita, Asier Mujika, Igor Petrovski, Pierre-Louis Cedoz, Chenmei Li, Shiyuan Chen, Niccolò Dal Santo, Siddharth Goyal, Jitesh Punjabi, Karthik Kappaganthu, Chester Kwak, Pallavi LV, Sarmishta Velury, Himadri Choudhury, Jamie Hall, Premal Shah, Ricardo Figueira, Matt Thomas, Minjie Lu, Ting Zhou, Chintu Kumar, Thomas Jurdi, Sharat Chikkerur, Yenai Ma, Adams Yu, Soo Kwak, Victor Ähdel, Sujeevan Rajayogam, Travis Choma, Fei Liu, Aditya Barua, Colin Ji, Ji Ho Park, Vincent Hellendoorn, Alex Bailey, Taylan Bilal, Huanjie Zhou, Mehrdad Khatir, Charles Sutton, Wojciech Rzadkowski, Fiona Macintosh, Roopali Vij, Konstantin Shagin, Paul Medina, Chen Liang, Jinjing Zhou, Pararth Shah, Yingying Bi, Attila Dankovics, Shipra Banga, Sabine Lehmann, Marissa Bredesen, Zifan Lin, John Eric Hoffmann, Jonathan Lai, Raynald Chung, Kai Yang, Nihal Balani, Arthur Bražinskas, Andrei Sozanschi, Matthew Hayes, Héctor Fernández Alcalde, Peter Makarov, Will Chen, Antonio Stella, Liselotte Snijders, Michael Mandl, Ante Kärrman, Paweł Nowak, Xinyi Wu, Alex Dyck, Krishnan Vaidyanathan, Raghavender R, Jessica Mallet, Mitch Rudominer, Eric Johnston, Sushil Mittal, Akhil Udathu, Janara Christensen, Vishal Verma, Zach Irving, Andreas Santucci, Gamaleldin Elsayed, Elnaz Davoodi, Marin Georgiev, Ian Tenney, Nan Hua, Geoffrey Cideron, Edouard Leurent, Mahmoud Alnahlawi, Ionut Georgescu, Nan Wei, Ivy Zheng, Dylan Scandinaro, Heinrich Jiang, Jasper Snoek, Mukund Sundararajan, Xuezhi Wang, Zack Ontiveros, Itay Karo, Jeremy Cole, Vinu Rajashekhar, Lara Tumeh, Eyal Ben-David, Rishub Jain, Jonathan Uesato, Romina Datta, Oskar Bunyan, Shimu Wu, John Zhang, Piotr Stanczyk, Ye Zhang, David Steiner, Subhajit Naskar, Michael Azzam, Matthew Johnson, Adam Paszke, Chung-Cheng Chiu, Jaume Sanchez Elias, Afroz Mohiuddin, Faizan Muhammad, Jin Miao, Andrew Lee, Nino Vieillard, Jane Park, Jiageng Zhang, Jeff Stanway, Drew Garmon, Abhijit Karmarkar, Zhe Dong, Jong Lee, Aviral Kumar, Luowei Zhou, Jonathan Evens, William Isaac, Geoffrey Irving, Edward Loper, Michael Fink, Isha Arkatkar, Nanxin Chen, Izhak Shafran, Ivan Petrychenko, Zhe Chen, Johnson Jia, Anselm Levskaya, Zhenkai Zhu, Peter Grabowski, Yu Mao, Alberto Magni, Kaisheng Yao, Javier Snaider, Norman Casagrande, Evan Palmer, Paul Suganthan, Alfonso Castaño, Irene Giannoumis, Wooyeol Kim, Mikołaj Rybiński, Ashwin Sreevatsa, Jennifer Prendki, David Soergel, Adrian Goedeckemeyer, Willi Gierke, Mohsen Jafari, Meenu Gaba, Jeremy Wiesner, Diana Gage Wright, Yawen Wei, Harsha Vashisht, Yana Kulizhskaya, Jay Hoover, Maigo Le, Lu Li, Chimezie Iwuanyanwu, Lu Liu, Kevin Ramirez, Andrey Khorlin, Albert Cui, Tian LIN, Marcus Wu, Ricardo Aguilar, Keith Pallo, Abhishek Chakladar, Ginger Perng, Elena Allica Abellan, Mingyang Zhang, Ishita Dasgupta, Nate Kushman, Ivo Penchev, Alena Repina, Xihui Wu, Tom van der Weide, Priya Ponnapalli, Caroline Kaplan, Jiri Simsa, Shuangfeng Li, Olivier Dousse, Fan Yang, Jeff Piper, Nathan Ie, Rama Pasumarthi, Nathan Lintz, Anitha Vijayakumar, Daniel Andor, Pedro Valenzuela, Minnie Lui, Cosmin Paduraru, Daiyi Peng, Katherine Lee, Shuyuan Zhang, Somer Greene, Duc Dung Nguyen, Paula Kurylowicz, Cassidy Hardin, Lucas Dixon, Lili Janzer, Kiam Choo, Ziqiang Feng, Biao Zhang, Achintya Singhal, Dayou Du, Dan McKinnon, Natasha Antropova, Tolga Bolukbasi, Orgad Keller, David Reid, Daniel Finchelstein, Maria Abi Raad, Remi Crocker, Peter Hawkins, Robert Dadashi, Colin Gaffney, Ken Franko, Anna Bulanova, Rémi Leblond, Shirley Chung, Harry Askham, Luis C. Cobo, Kelvin Xu, Felix Fischer, Jun Xu, Christina Sorokin, Chris Alberti, Chu-Cheng Lin, Colin Evans, Alek Dimitriev, Hannah Forbes, Dylan Banarse, Zora Tung, Mark Omernick, Colton Bishop, Rachel Sterneck, Rohan Jain, Jiawei Xia, Ehsan Amid, Francesco Piccinno, Xingyu Wang, Praseem Banzal, Daniel J. Mankowitz, Alex Polozov, Victoria Krakovna, Sasha Brown, MohammadHossein Bateni, Dennis Duan, Vlad Firoiu, Meghana Thotakuri, Tom Natan, Matthieu Geist, Ser tan Girgin, Hui Li, Jiayu Ye, Ofir Roval, Reiko Tojo, Michael Kwong, James Lee-Thorp, Christopher Yew, Danila Sinopalnikov, Sabela Ramos, John Mellor, Abhishek Sharma, Kathy Wu, David Miller, Nicolas Sonnerat, Denis Vnukov, Rory Greig, Jennifer Beattie, Emily Caveness, Libin Bai, Julian Eisenschlos, Alex Korchemniy, Tomy Tsai, Mimi Jasarevic, Weize Kong, Phuong Dao, Zeyu Zheng, Frederick Liu, Fan Yang, Rui Zhu, Tian Huey Teh, Jason Sanmiya, Evgeny Gladchenko, Nejc Trdin, Daniel Toyama, Evan Rosen, Sasan Tavakkol, Linting Xue, Chen Elkind, Oliver Woodman, John Carpenter, George Papamakarios, Rupert Kemp, Sushant Kafle, Tanya Grunina, Rishika Sinha, Alice Talbert, Diane Wu, Denese Owusu-Afriyie, Cosmo Du, Chloe Thornton, Jordi Pont-Tuset, Pradyumna Narayana, Jing Li, Saaber Fatehi, John Wieting, Omar Ajmeri, Benigno Uria, Yeongil Ko, Laura Knight, Amélie Héliou, Ning Niu, Shane Gu, Chenxi Pang, Yeqing Li, Nir Levine, Ariel Stolovich, Rebeca Santamaria-Fernandez, Sonam Goenka, Wenny Yustalim, Robin Strudel, Ali Elqursh, Charlie Deck, Hyo Lee, Zonglin Li, Kyle Levin, Raphael Hoffmann, Dan Holtmann-Rice, Olivier Bachem, Sho Arora, Christy Koh, Soheil Hassas Yeganeh, Siim Põder, Mukarram Tariq, Yanhua Sun, Lucian Ionita, Mojtaba Seyedhosseini, Pouya Tafti, Zhiyu Liu, Anmol Gulati, Jasmine Liu, Xinyu Ye, Bart Chrzaszcz, Lily Wang, Nikhil Sethi, Tianrun Li, Ben Brown, Shreya Singh, Wei Fan, Aaron Parisi, Joe Stanton, Vinod Koverkathu, Christopher A. Choquette-Choo, Yunjie Li, TJ Lu, Abe Ittycheriah, Prakash Shroff, Mani Varadarajan, Sanaz Bahargam, Rob Willoughby, David Gaddy, Guillaume Desjardins, Marco Cornero, Brona Robenek, Bhavishya Mittal, Ben Albrecht, Ashish Shenoy, Fedor Moiseev, Henrik Jacobsson, Alireza Ghaffarkhah, Morgane Rivière, Alanna Walton, Clément Crepy, Alicia Parrish, Zongwei Zhou, Clement Farabet, Carey Radebaugh, Praveen Srinivasan, Claudia van der Salm, Andreas Fidjeland, Salvatore Scellato, Eri Latorre-Chimoto, Hanna Klimczak-Plucińska, David Bridson, Dario de Cesare, Tom Hudson, Piermaria Mendolicchio, Lexi Walker, Alex Morris, Matthew Mauger, Alexey Guseynov, Alison Reid, Seth Odoom, Lucia Loher, Victor Cotruta, Madhavi Yenugula, Dominik Grewe, Anastasia Petrushkina, Tom Duerig, Antonio Sanchez, Steve Yadlowsky, Amy Shen, Amir Globerson, Lynette Webb, Sahil Dua, Dong Li, Surya Bhupatiraju, Dan Hurt, Haroon Qureshi, Ananth Agarwal, Tomer Shani, Matan Eyal, Anuj Khare, Shreyas Rammohan Belle, Lei Wang, Chetan Tekur, Mihir Sanjay Kale, Jinliang Wei, Ruoxin Sang, Brennan Saeta, Tyler Liechty, Yi Sun, Yao Zhao, Stephan Lee, Pandu Nayak, Doug Fritz, Manish Reddy Vuyyuru, John Aslanides, Nidhi Vyas, Martin Wicke, Xiao Ma, Evgenii Eltyshev, Nina Martin, Hardie Cate, James Manyika, Keyvan Amiri, Yelin Kim, Xi Xiong, Kai Kang, Florian Luisier, Nilesh Tripuraneni, David Madras, Mandy Guo, Austin Waters, Oliver Wang, Joshua Ainslie, Jason Baldridge, Han Zhang, Garima Pruthi, Jakob Bauer, Feng Yang, Riham Mansour, Jason Gelman, Yang Xu, George Polovets, Ji Liu, Honglong Cai, Warren Chen, XiangHai Sheng, Emily Xue, Sherjil Ozair, Christof Angermueller, Xiaowei Li, Anoop Sinha, Weiren Wang, Julia Wiesinger, Emmanouil Koukoumidis, Yuan Tian, Anand Iyer, Madhu Gurumurthy, Mark Goldenson, Parashar Shah, MK Blake, Hongkun Yu, Anthony Urbanowicz, Jennimaria Palomaki, Chrisantha Fernando, Ken Durden, Harsh Mehta, Nikola Momchev, Elahe Rahimtoroghi, Maria Georgaki, Amit Raul, Sebastian Ruder, Morgan Redshaw, Jinhyuk Lee, Denny Zhou, Komal Jalan, Dinghua Li, Blake Hechtman, Parker Schuh, Milad Nasr, Kieran Milan, Vladimir Mikulik, Juliana Franco, Tim Green, Nam Nguyen, Joe Kelley, Aroma Mahendru, Andrea Hu, Joshua Howland, Ben Vargas, Jeffrey Hui, Kshitij Bansal, Vikram Rao, Rakesh Ghiya, Emma Wang, Ke Ye, Jean Michel Sarr, Melanie Moranski Preston, Madeleine Elish, Steve Li, Aakash Kaku, Jigar Gupta, Ice Pasupat, Da-Cheng Juan, Milan Someswar, Tejvi M., Xinyun Chen, Aida Amini, Alex Fabrikant, Eric Chu, Xuanyi Dong, Amruta Muthal, Senaka Buthpitiya, Sarthak Jauhari, Nan Hua, Urvashi Khandelwal, Ayal Hitron, Jie Ren, Larissa Rinaldi, Shahar Drath, Avigail Dabush, Nan-Jiang Jiang, Harshal Godhia, Uli Sachs, Anthony Chen, Yicheng Fan, Hagai Taitelbaum, Hila Noga, Zhuyun Dai, James Wang, Chen Liang, Jenny Hamer, Chun-Sung Ferng, Chenel Elkind, Aviel Atias, Paulina Lee, Vít Listík, Mathias Carlen, Jan van de Kerkhof, Marcin Pikus, Krunoslav Zaher, Paul Müller, Sasha Zykova, Richard Stefanec, Vitaly Gatsko, Christoph Hirnschall, Ashwin Sethi, Xingyu Federico Xu, Chetan Ahuja, Beth Tsai, Anca Stefanoiu, Bo Feng, Keshav Dhandhania, Manish Katyal, Akshay Gupta, Atharva Parulekar, Divya Pitta, Jing Zhao, Vivaan Bhatia, Yashodha Bhavnani, Omar Alhadlaq, Xiaolin Li, Peter Danenberg, Dennis Tu, Alex Pine, Vera Filippova, Abhipso Ghosh, Ben Limonchik, Bhargava Urala, Chaitanya Krishna Lanka, Derik Clive, Yi Sun, Edward Li, Hao Wu, Kevin Hongtongsak, Ianna Li, Kalind Thakkar, Kuanysh Omarov, Kushal Majmundar, Michael Alverson, Michael Kucharski, Mohak Patel, Mudit Jain, Maksim Zabelin, Paolo Pelagatti, Rohan Kohli, Saurabh Kumar, Joseph Kim, Swetha Sankar, Vineet Shah, Lakshmi Ramachandruni, Xiangkai Zeng, Ben Bariach, Laura Weidinger, Tu Vu, Alek Andreev, Antoine He, Kevin Hui, Sheleem Kashem, Amar Subramanya, Sissie Hsiao, Demis Hassabis, Koray Kavukcuoglu, Adam Sadovsky, Quoc Le, Trevor Strohman, Yonghui Wu, Slav Petrov, Jeffrey Dean, Oriol Vinyals (arXiv:2312.11805), used under CC BY 4.0. Changes: condensed and restructured for implementation, with tables, equations and algorithms taken from the paper and its code, each cited to its source. Not affiliated with or endorsed by the authors.