The State of
Machine Learning Competitions

ML Competitions Landscape

More than 390 machine learning competitions took place in 2025, across 30+ platforms, with a total prize pool of over $16m. These competitions included multi-million-dollar government-funded competitions, grand challenges like the AI Mathematical Olympiad, and research competitions at ML or computer vision conferences.¹

This report starts with a brief overview of the landscape of competition platforms. For a dissection of the approaches used by competition winners in 2025, skip ahead to winning solutions.

Platforms

Kaggle remains by far the platform with the most users; it ran the most competitions in 2025, as well as having the largest total prize pool.² Tianchi is the second-largest platform by all three of these metrics. Codabench hosted the third-most competitions in 2025, and its user count more than doubled over the year, a year in which multiple platforms saw 30%+ user growth and Zindi became the sixth platform to have over 100,000 users.

2025 Platform Comparison

Overview of Top Competition Platforms

The open-source platform CodaLab is being sunset, and fully replaced by its newer sibling Codabench.⁹ Codabench is also open-source, developed by the same team, who host a free-to-use instance of the platform.

The other main open-source self-service competitions platform, EvalAI, appears to be introducing hosting fees for the first time, with its website stating it is “transitioning to a paid plan for hosting challenges”.¹⁰

At the same time, Kaggle’s free-to-use self-service “Community Competitions” now allow organisers to offer prizes up to $10k. Of the 68 Kaggle competitions in our 2025 dataset, 34 were Community competitions.

As well as growing by another seven million users, Kaggle’s product saw more significant features added in 2025 than in previous years. New features include Benchmarks (ongoing leaderboards aimed at comparing frontier models’ performance on standardised tasks), Game Arena (player-vs-player tournaments, pitting models against each other in settings like chess, poker, or Werewolf), and Writeups (long-form blog posts, including for competition solutions). Kaggle’s progression system — which sets incentives and determines the criteria for becoming a Kaggle “Master” or “Grandmaster” — also saw a significant overhaul in 2025.

Other platforms also added features in 2025 — Zindi added multi-metric evaluation and new types of competitions, while the open-source Codabench added numerous quality-of-life features for competition participants and hosts.

Hugging Face’s Competitions product had its last commit in October 2024, and the product lead left the company at the beginning of 2025. It still hosted two competitions in 2025, with a total prize pool of $16k.

DrivenData published a review of what happens after competitions have concluded, including several case studies detailing real-world systems deployment, research publications, and hiring decisions based on competition submissions.

Prizes & Participation

The largest prize pool in 2025 was $8.5m, for the final round of DARPA’s AI Cyber Challenge.

The number of competitions with prize pools of $100k or greater has grown consistently in the past four years; we found 23 such competitions in 2025, up from 18 in 2024.

While the prize pool distribution is usually skewed towards first place (with the first-placed winner awarded the largest portion of the pool), this year, some judging-panel-evaluated competitions awarded equal prizes to each of the winners. For example, in the OpenAI gpt-oss-20b red-teaming challenge, each of the top ten teams received a $50k prize.

Prize pool

In general, the competitions with larger prize pools drew more competitors. Aside from the obvious direct draw of the prize money, competitions with large prizes — especially those posed as ‘grand challenges’ with $1m+ prize pools conditional on absolute performance — may be embraced as significant milestones by the broader machine learning community. One example of this is the ARC Prize — since its million-dollar grand prize relaunch in 2024,¹¹ its public leaderboard is now often used as a yardstick for reasoning performance on new frontier models,¹² as well as being a core research topic for startups (including Tufa Labs, Poetiq, Basis, Agemo and Giotto.ai).

The second AIMO progress prize had a guaranteed prize pool of over half a million dollars, with an additional $1.6m available conditional on performance, and drew entries from over 2,200 teams. The CIBMTR hematopoietic cell transplantation survival rate prediction competition, with a $50k prize pool, drew submissions from over 3,000 teams.

Direct community engagement can also attract a large pool of qualified competitors. The GPU Mode Discord community, which started as a reading group for a parallel programming book in 2024, has since grown to over 24,000 members, and hosted two $150k AMD GPU kernel inference competitions in 2025. The first competition received submissions from more than 160 teams, a notable amount for an independently-hosted competition. Alongside competitions, GPU Mode also hosts a lecture series related to kernel development and large language models.

Leaderboard entries

The most popular competition we found in 2025 was Jane Street’s Real-Time Market Data Forecasting competition, with over 3,700 teams taking part and a $120k prize pool. The median number of teams per competition in our dataset is 36, and the mean is 351.

Many of the competitions at the lower end of this scale are part of competition tracks at conferences, which are often targeted at researchers within a specific niche rather than the general competitions community.

Winning Solutions

To track tools and techniques used by competition winners, we analysed the #1-ranked solution for as many of the competitions in 2025 as we could,¹³ and gathered information on programming languages, packages, model architectures, and computational resources.

Almost all winning teams used Python as the main programming language for their solutions, in line with previous years. The one exception we found was Kaggle’s Efficient Chess competition, for which the winner used C as their primary language, alongside Rust and Python.

The packages listed below represent the key Python packages that make up competition winners’ toolkit.

Python Packages

New Additions: Highlights

librosa is a package for music and audio analysis. It was used in several audio-focused competitions, mainly for its audio transforms including normalisation and pitch shifting.

monai, the Medical Open Network for Artificial Intelligence, provides a framework for deep learning in healthcare imaging, built on top of PyTorch. It was used by the winners of several healthcare-related competitions, including the team that won Kaggle’s CryoET Object Identification competition.

torchmetrics provides implementations of over 100 metrics in PyTorch.

onnx is an open format for machine learning interoperability, and onnxruntime is a cross-platform inference and training accelerator, developed by Microsoft. Models can be converted to the ONNX standard from many other frameworks (including PyTorch). Winning solutions using the ONNX Runtime mostly used it to make model inference more efficient.

hydra is a configuration management framework. Working on ML competitions often involves running many experiments, and hydra is designed to make it easy to maintain configuration and manage overrides, whether through configuration files or command line arguments.

unsloth is a library focused on efficient LLM fine-tuning. Built on top of PyTorch, it enables faster training with lower memory use through a variety of optimisations, including custom GPU kernels. It can also be used to speed up inference.

vLLM is a library for LLM inference and serving, including features like batching, quantisation, speculative decoding, and optimised cuda kernels.

autogluon is an AutoML library that aims to build powerful ensembles, making use of many other modelling libraries, and combines these with bagging and stack-ensembling to maximise predictive power without manual hyperparameter tuning. It was used for tabular prediction in two winning solutions.

Deep Learning

Deep Learning Libraries

PyTorch remains the dominant deep learning framework; we found 44 winning solutions using PyTorch and only 1 using TensorFlow.

Of the solutions using PyTorch, 9 also used the higher-level PyTorch Lightning library, which provides additional ready-to-use wrappers for training loops, data preparation, and other components. For examples of how PyTorch Lightning was used, see the winning solutions for Zindi’s Côte d’Ivoire Byte-Sized Agriculture Challenge or the Vesuvius First Title Prize.

The one winning solution we found using TensorFlow was for the Geology Forecast Challenge, a community competition on Kaggle, where the winner used TensorFlow’s higher-level Keras API.

Despite its popularity in other domains¹⁴ we didn’t find any winners using JAX to design or train their models. Even for the Lux AI Season 3 competition, where the environment was implemented in JAX, the winning solution used PyTorch.

One winning solution used the Rust-based bullet framework to train a small neural net for evaluating chess positions.

DataFrames

DataFrames remain a two-bear race: 61 of the analysed winning solutions used Pandas, and 5 used Polars.

Of the 5 that used Polars, two used it minimally for IO, two used a mix of Polars and Pandas (e.g. some team members used Pandas, some Polars), and one used Polars as its primary DataFrame library with Pandas only for interfacing with other machine learning libraries.

For more analysis of Polars use in solutions, see the 2024 and 2023 editions of this report.

Tabular Data

Gradient-boosted decision trees (GBDTs) remain the go-to tabular competition winner’s modelling tool, sometimes as part of an ensemble alongside neural nets.

Among winning solutions that used GBDTs, XGBoost and LightGBM were most popular this year, with 14 uses each. We found 8 winning solutions using CatBoost. Although part of the choice of library is likely due to familiarity, some winners explicitly stated that they experimented with all three of these libraries, and found significant performance differences between the different libraries on their dataset.¹⁵

AutoGluon was used for tabular prediction in two winning solutions. Most notably, the winner of Kaggle’s Open Polymer Prediction 2025 competition used AutoGluon as part of an ensemble, and noted that AutoGluon “was able to beat an ensemble of XGBoost, LightGBM, and TabM models [tuned with] Optuna”, while using only a fraction of the training budget.¹⁶

For the first time, we also saw a tabular foundation model (TabPFN) used, by the winners of DrivenData’s PREPARE challenge. The winning team used TabPFN to generate features that were fed into GBDT models.

We also found one winning solution using the TabM tabular deep learning model to predict transplant survival rates in a Kaggle competition. Their solution used TabM as part of an ensemble of models, including GBDTs and a combination of k-nearest-neighbours and graph neural nets.

Computer Vision

Computer Vision Architectures

For the first time, Transformer-based architectures outnumbered those based on convolutional neural networks (CNNs) among winning solutions.

The most popular Transformer-based architectures included ViT (often DinoV2) and Swin Transformer for image classification, regression, or segmentation tasks, and Qwen-VL models for image-to-text tasks.

The most popular CNN-based architectures included YOLOv8 and YOLO11 for object detection, and ResNet and EfficientNet for other tasks.

Audio

Audio competitions in 2025 included tasks related to speech recognition, animal species identification from audio clips, student literacy assessment from audio clips, and dementia screening using acoustic biomarkers from voice recordings.

In four of the five competitions where the audio included human speech, the winners fine-tuned versions of OpenAI’s Transformer-based Whisper model. The fifth used Meta’s W2v-BERT 2.0, based on Conformers, a Transformer/CNN hybrid architecture.

The winner of the BirdCLEF+ 2025 species identification competition used a mixture of CNN-based models, each with sound-event-detection heads based on those in a top solution to the 2021 edition of this competition.

Natural Language Processing

The ongoing shift from encoder models (like DeBERTa) to decoder models (generative LLMs) like Llama/Mistral/Gemma/Qwen/DeepSeek continued in 2025.¹⁷

Alibaba’s Qwen was the clear winner among open LLMs in 2025. Almost all of the winning solutions we found for competitions with a text-processing element involved either a Qwen2.5 model, a Qwen3 model, or both. We found limited use among winners of other open LLM models.

The winners of all three of the Kaggle competitions with large conditional grand prizes used Qwen models: the winner of the Konwinski prize used Qwen2.5-Coder-32B, the winner of the AIMO progress prize 2 used Qwen2.5-14B, and the winner of ARC Prize 2025 used Qwen3-4B (as well as other models for synthetic data generation).

We found only one winning team that made use of a BERT-style encoder model. The solution to the Open Polymer Prediction competition made use of both CodeBERT and ModernBERT, with full fine-tuning.

Weight quantisation was common among winning NLP solutions, particularly at inference time. Partial fine-tuning using LoRA was also common, though some teams opted to do full fine-tuning, requiring larger compute budgets. See last year’s report for more detail on LoRA and quantisation.

We found three winning solutions using Unsloth for efficient fine-tuning and inference of transformer-based neural nets implemented in PyTorch. Two of these used test-time fine-tuning, where the compute constraints of the evaluation environment made efficiency essential. Four winning solutions used vLLM for efficient LLM inference.

Compute and Hardware

There was a shift towards more compute-intensive solutions this year, and we found at least three instances where multi-node corporate or university-owned clusters were used to train competition-winning solutions.

Nonetheless, there were winning teams at the other end of the spectrum, using (mostly free) cloud notebook services for training. Nine used Kaggle Notebooks, one used Google Colab’s Free tier, and one used Google Colab’s Pro tier. Kaggle’s Notebooks generally come with older GPU or TPU hardware, though certain recent Kaggle competitions have allowed participants to use newer L4 or H100 GPUs.

We also found six winners using personal hardware, and three using rented cloud servers. For the majority of winning solutions we analysed, we were unable to identify whether the training hardware was owned or rented.

Hardware Used by Winners

NVIDIA’s H100 finally dethroned the older A100 as the most-commonly-used GPU among winning solutions. All the winning solutions we found were trained either on NVIDIA GPUs or exclusively on CPUs.¹⁸ We did not find any winning teams using AMD GPUs or Google TPUs for training.

Accelerator Models

Along with the H100, an NVIDIA Hopper generation GPU, first released in 2022, we also saw two of the new NVIDIA Blackwell generation GPU models being used for training winning solutions: the RTX 5090 and the RTX Pro 6000 Blackwell, both released in 2025.

Some winning solutions were trained on GPUs released almost a decade ago. In particular, the NVIDIA T4 and P100, which are both still available in Kaggle’s notebook environments, were used to train several winning solutions.

Multi-node setups

For the first time this year, we saw several multi-node training setups. The winners of the WSDM Cup and ARC Prize 2025 both used four 8xH100 nodes for training, a team from Meta used 128 of the older V100 GPUs for their winning solution to the Algonauts Challenge, and the winners of the second AI Mathematical Olympiad Progress Prize used 512 H100 GPUs for training. To quote their write-up, “training lasted 48 hours on 512 H100 (yes, 512!)”.

Training Budgets

Budgets

While the majority of winning solutions would have cost under $100 to train using on-demand cloud compute,¹⁹ some winning solutions used much greater computational resources.

At the high end, 512 H100s were used for 48 hours of training,²⁰ with an estimated on-demand cost of over $60k.²¹ The second-highest (with a cost estimate of just over $2.5k) used 32 H100s for 27 hours and 8 H100s for 24 hours,²² for ARC Prize 2025. Both of these teams were made up entirely of NVIDIA employees.²³ These training budgets are much larger than those used for competition-winning solutions in previous years (although still several orders of magnitude below the training budgets for frontier models).

The largest training budget we found in 2024 was around $500, and in 2025 at least four competition winners exceeded that, including the two NVIDIA teams mentioned above.

Top Solutions

This section highlights a few of 2025’s most innovative or widely-applicable winning solutions.

Efficient Chess AI Challenge (Linmiao Xu)

This competition required building a chess agent under tight CPU and memory constraints. Xu’s winning solution paired the Cfish chess engine with a neural net for position evaluation, trained using the Rust-based chess neural net framework bullet. The neural net used an NNUE architecture; a type of incrementally-updatable neural net developed for sparse binary input spaces with minimal input changes between inference passes.²⁴ This approach, which also used carefully-considered quantisation and compression, beat over 1,000 other teams in a highly resource-constrained environment.

-> Read the full solution write-up here

Agile Community Rules Classification (Guanshuo Xu)

The Jigsaw Agile Community Rules Classification competition had an unusual feature: each data row included positive and negative examples — comments which do/do not violate the given moderation rule — as additional inputs that could be used by competitors’ models.²⁵ Xu’s winning solution generated additional training data from these examples and used that for fine-tuning — all within the Kaggle competition evaluation environment intended for inference, and within the 12-hour time limit. Using an ensemble of 6 LLMs (with up to 14 billion parameters each) within these constraints required switching from PyTorch to Unsloth, a framework for fast, efficient model training and inference, built on top of PyTorch, as well as a few other optimisations for both training and inference.

-> Read the full solution write-up here

Geophysical Waveform Inversion (Harshit Sheoran)

This was a “prediction” competition, where participants submit their predictions for the test data, as opposed to a (now more common) “code” competition, where participants submit code that will generate predictions on the private test set. The core task in the competition: given the output (seismic data) of a physics simulation, predict the input (subsurface velocity maps).

Sheoran’s success on this problem came both from improvements in model architecture and creating a positive feedback loop between data generation and model training.

After building a model to predict the simulation input, Sheoran iteratively trained this model by running the simulation forward to get the matching output for each ‘predicted input’. These newly-labelled (simulation output, simulation input) pairs were then added to the training set for the next epoch, and the process would repeat.

-> Read the full solution write-up here

ARC Prize 2025 (Team NVARC)

In the second iteration of the annual ARC Prize competition, NVIDIA’s Team NVARC built on the winning solution to the previous iteration (Franzen et al., 2025), as well as the Tiny Relational Model (TRM) architecture that won the best paper award (Jolicoeur-Martineau, 2025), and combined that with a multi-stage synthetic data generation pipeline with hundreds of thousands of synthetic samples. They achieved a score of 24.03 on the private test set; a lead of almost 50% over second-place’s 16.53. Their write-up also describes the outcome of some experiments they conducted after the end of the competition.

-> Read the full solution write-up here

AI Mathematical Olympiad Progress Prize 2 (Team NemoSkills)

The winning solution to the second AIMO progress prize used large-scale synthetic data, tool-integrated reasoning, model merging, significant inference optimisations, and the largest compute budget of any winning solution in 2025. The team published their synthetic training data, as well as a paper (Moshkov et al., 2025) describing their approach in more detail.

-> Read the full solution write-up here

Team Demographics

More than half of the winning teams we found in 2025 were made up of just one person, for the third year in a row. Teams of more than five were rare.

Winning Team Sizes

Over half of the winning teams we found were categorised as ‘first-time winners’, with no members who had already won a competition on the same platform.

Repeat Winners

As mentioned in the previous sections, two of the most significant Kaggle competitions of 2025 — ARC Prize 2025 and AIMO Progress Prize 2 — were won by teams of NVIDIA employees, including members of their “KGMON” Kaggle Grandmasters Of NVIDIA group. This group makes up more than 10% of the current top-50-ranked Kaggle users.²⁶

Some of the members of this group published an article titled The Kaggle Grandmasters Playbook in September 2025, giving an accessible introduction to seven techniques, including stacking and pseudo-labelling.

Trends

Academia

Kaggle published a position paper at the International Conference on Machine Learning (ICML), making the case for competitions as the gold standard for empirical rigour in Generative AI evaluation.

NeurIPS

The 2025 conference on Neural Information Processing Systems, NeurIPS, marked the tenth year of its official competitions track. In this edition, the conference accepted 18 official competitions, with a total prize pool of $270k. Four of the competitions were hosted on Kaggle, three on Codabench, and three on EvalAI.

The NeurIPS 2025 competition with the most prize money was the Google Code Golf Championship, hosted on Kaggle, which had a $100k prize pool and attracted over 1,100 teams. The competition involved writing the shortest possible programs to solve ARC Prize problems from the ARC-AGI-1 dataset. Other NeurIPS 2025 competitions included problems related to weather forecasting, exoplanet signal extraction, robotics and control, and various aspects of LLMs.

NeurIPS Competitions

Other conferences

Most of the competitions at the International Conference on Medical Image Computing and Computer Assisted Intervention (MICCAI) were hosted on the Grand Challenge platform. Computer vision workshop competitions at the International Conference on Computer Vision (ICCV) and the Conference on Computer Vision and Pattern Recognition (CVPR) were mostly hosted on CodaLab or EvalAI.²⁷

Multi-Year Competitions

Some updates on ambitious competitions with multi-year agendas. Many of these have headline grand prizes, often $1m+, paid out when a certain performance milestone is achieved, in addition to guaranteed prizes given out to top performing teams in each iteration of the competition.

In the second progress prize of the AI Mathematical Olympiad (AIMO), the winning team took home just over a quarter of a million dollars. With a score of 34/50 on the private test set, they improved on the high score of 29 in the first progress prize, despite significantly increased problem difficulty, but they were still a long way short of the 47/50 needed to unlock the $1.5m+ grand prize. The third AIMO progress prize, with harder problems and a similar conditional prize structure, is currently ongoing. It ends in April 2026, and at the time of writing nine teams on the public leaderboard have a score of 44/50 — just three points short of the grand prize level.²⁸

The finals for DARPA’s AI Cyber Challenge took place at DEF CON 33 in August, with Team Atlanta taking home the $4m prize for first place. Their write-up can be found on their website.

In the first edition of the Konwinski Prize, the winner took home a $50k prize for resolving 8% of the GitHub issues in the test set. Additional prizes would have been unlocked at 10% increments, starting from 30%, up to the 90% success rate required for the $800k+ top prize. The winner, who had a significant lead over others on the leaderboard, credits context management as one of the most important components of their solution. The top five teams’ solutions all used either Qwen or DeepSeek models in their solution.

The Vesuvius Challenge paid out close to $100k across multiple prizes in 2025, most notably the $60k first title prize for reading the title and author of one of the scrolls non-invasively. Other progress on their research agenda in 2025 included the first letters in scroll 4 being read.

ARC Prize 2025 concluded with the leading solution reaching 24% on the ARC-AGI-2 private evaluation set, at a cost of approximately $0.20/task. At the time of writing, the best closed model with a similar cost (Gemini 3 Flash Preview) achieves 33.6%, and the overall best (closed model) solution achieves 84.6% with a cost of around $14/task.²⁹ The winning team took home $25k. A score of 85% was needed to unlock the additional $700k grand prize.

After the conclusion of ARC Prize 2025, the ARC Prize team published a blog post and comprehensive technical report with a summary of the results and approaches taken by successful teams. The organisers identified “refinement loops” as the defining theme of 2025, stating that methods incorporating weight-space refinement loops “are now achieving competitive performance with remarkably small networks (7M parameters).”

The winning solution incorporated a Tiny Reasoning Model, as described in the paper which won ARC Prize 2025’s best paper award and received significant attention from the broader machine learning community (Jolicoeur-Martineau, 2025).

The organisers aim to make it harder for systems to overfit to ARC problems in the third version of the ARC dataset, ARC-AGI-3, the first to feature dynamic game-like environments, with ARC Prize 2026 launching March 25, 2026.³⁰

Looking Ahead

Despite the previous claims of “Kaggle grandmaster-level” autonomous agents, noted last year, we have yet to see any publications demonstrating successful use of these agents in live competitions. The advent of improved code generation models and harnesses, as well as evolutionary approaches like AlphaEvolve (Novikov et al., 2025), may accelerate progress in this area.

About This Report

About ML Contests

For over six years now, ML Contests has provided a competition directory and shared insights on trends in machine learning competitions. To receive occasional updates with content like this report, subscribe to our mailing list or RSS feed.

If you enjoyed reading this report, you might like the previous editions.

About Jolt

ML Contests is part of Jolt, an online magazine for ML researchers and practitioners.

Jolt publishes long-read technical articles, as well as updates and reporting from major machine learning conferences.

Acknowledgements

Thank you to Mia Van Petegem, Peter Carlens, Alex Cross, and James Mundy, for helpful feedback on drafts of this report.

Thank you to the teams at AIcrowd, Antigranular, CrunchDAO, DrivenData, Grand Challenge, Kaggle, Solafune, ThinkOnward, Trustii, and Zindi, for providing data on their competitions and winning solutions.

Thank you to the competition winners who took the time to answer our questions by email or questionnaire: Kea Kohv, Harshit Sheoran, James Day, Darius Moruri, TonyK, Levi Harris, Chigozie Nkwocha, Julius Mwangi M., Benson Wainaina, Duke Kojo Kongo, and Naoto Suzuki.

Lastly, thank you to the maintainers of the open source projects we made use of in conducting our research and producing this page: Astro, Chart.js, Linguist, pipreqs, and nbconvert.

Methodology

Data was gathered from correspondence with competition platforms, organisers, and winners, as well as from public online sources.

The following criteria were applied for inclusion. Each competition must:

• have a total available prize pool of at least $1,000 in cash or liquid cryptocurrency, or be an official competition at a top machine learning or robotics conference with a dedicated competitions track (such as NeurIPS, MICCAI, or ICRA)
• have ended between 1 January 2025 and 31 December 2025 (inclusive).

Where available, we used data provided by each competition platform, describing their own competitions, as a starting point. Ten platforms were able to provide us with this data.

We did not receive this data from all platforms, and in these cases we gathered the data from their website. Notably, this applied to CodaLab, Codabench, and EvalAI (three ‘self-service’ platforms where competition organisers can create their own competitions with minimal to no intervention from the platform team). For these platforms, we did our best to collect complete data and filter out irrelevant competitions (such as ones used for class assignments or draft competitions which never ended up running).

Our filter for conference-related competitions was narrower than in previous years. In previous years, any competition affiliated with a machine learning or robotics conference was automatically eligible (including those at competition workshops). This resulted in us including a large number of competitions which had very few submissions, and about which public information was limited and hard to track down. This year we only included conference-related competitions if they were part of an official competitions track, and only those at “top conferences” — i.e. those with an A or A* ICORE ranking. Other conference-related competitions which meet the $1k minimum prize pool threshold were still included.

We were not able to collect full data for all platforms, especially those — like Tianchi and Signate — where English is not the primary language of the platform. It is likely that there are additional competition platforms as yet unknown to us. Please contact us if you know of any that are missing.

When counting a “number of competitions” for purposes such as prize pool distribution, or popularity of programming languages, we generally use the following definition: If a competition is made up of several tracks, each with separate leaderboards and separate prize pools, then each track counts as its own competition. If a competition is made up of multiple sub-tasks which are all measured together on one main leaderboard for one prize pool, they count together as one competition. There are some exceptions.⁶

For the purposes of this report, we consider a “competition winner” to be the #1-placed team in a competition as defined above. We are aware that other valid usages of the term exist, with their own advantages — for example, anyone winning a Gold/Silver/Bronze medal, or anyone winning a prize in a competition. For ease of analysis and in order to avoid double-counting, we exclusively consider #1-placed teams in this report when aggregating statistics on “winning solutions” or “competition winners”, and make occasional reference to other solutions.

Compiling the Python packages section in the winning toolkit involved some editorial choices. In general, we list packages which were either used in several winning solutions, or instrumental in the success of at least one winning solution.

The number of users for each platform is sourced from the platform directly, where possible. Some platforms list their user numbers publicly on their website. Most other platforms shared their user numbers with us via email. Solafune chose not to disclose their user numbers. User numbers for Signate, Tianchi, and Bitgrit are more than a year old, as we did not receive updates on their user numbers before publication of this report and were unable to find more recent public information.

[
    ('numpy', 62),
    ('pandas', 61),
    ('torch', 44),
    ('tqdm', 43),
    ('matplotlib', 34),
    ('sklearn', 33),
    ('scipy', 32),
    ('transformers', 25),
    ('torchvision', 16),
    ('seaborn', 16),
    ('cv2', 15),
    ('pil', 15),
    ('xgboost', 14),
    ('lightgbm', 14),
    ('timm', 13),
    ('requests', 12),
    ('yaml', 12),
    ('wandb', 11),
    ('huggingface_hub', 9),
    ('datasets', 9),
    ('pytorch_lightning', 9),
    ('peft', 9),
    ('catboost', 8),
    ('joblib', 8),
    ('setuptools', 8),
    ('albumentations', 8),
    ('skimage', 7),
    ('ipython', 7),
    ('einops', 6),
    ('librosa', 6),
    ('optuna', 6),
    ('monai', 6),
    ('onnxruntime', 6),
    ('rich', 6),
    ('torchmetrics', 6),
    ('polars', 6),
    ('simpleitk', 5),
    ('omegaconf', 5),
    ('pydantic', 5),
    ('accelerate', 4),
    ('vllm', 4),
    ('onnx', 4),
    ('mmengine', 4),
    ('safetensors', 4),
    ('xformers', 4),
    ('shap', 4),
    ('kagglehub', 4),
    ('evaluate', 4),
    ('packaging', 4),
    ('hydra', 4),
    ('soundfile', 3),
    ('pkg_resources', 3),
    ('click', 3),
    ('nibabel', 3),
    ('psutil', 3),
    ('mmcv', 3),
    ('plotly', 3),
    ('shapely', 3),
    ('numba', 3),
    ('geopandas', 3),
    ('tabulate', 3),
    ('bitsandbytes', 3),
    ('torchaudio', 3),
    ('unsloth', 3),
    ('tensorrt', 3),
    ('loguru', 3),
    ('jiwer', 3),
    ('pycocotools', 3)
]

Carlens, H, "State of Machine Learning Competitions in 2025", ML Contests, 2026.

@article{
carlens2026state,
author = {Carlens, Harald},
title = {State of Machine Learning Competitions in 2025},
journal = {ML Contests},
year = {2026},
note = {https://mlcontests.com/state-of-machine-learning-competitions-2025},
}