Open your canvas
OUR POSITION / PROGRESS BELONGS IN THE OPEN

Open Models

We believe great AI should be something more people can build on. For that reason, Underleau proudly supports and mostly uses open-weight models and champions the brilliant minds moving them forward.

Explore the models Why we choose open
01 / OPEN BY CONVICTION

A choice. Not a compromise.

Cost matters to us. So do capability, independence and the community that makes this work possible. We do not see open weights as a second-class foundation.

More people can build.

Published weights let more people inspect, adapt and build on a model, within its licence.

Efficiency is a feature.

Match the model to the work. Useful AI can be more economical, even though open weights still need compute.

Choice keeps things open.

Different models, tools and hosting options give builders room to improve and change direction.

02 / AN ECOSYSTEM, NOT A GATEKEEPER

Meet the open model community.

Well-known families, different strengths, one important idea: powerful models can be made available for others to build on.

04

Meta

Llama

Multimodal & long-context work

Meta's open-weight family, backed by a broad developer ecosystem.

Release & licence
Featured release
Llama 4 Scout
Licence
Llama 4 Community Licence

The Llama 4 community licence includes additional conditions and regional restrictions.

Read the original release (opens in a new tab)
05

Google DeepMind

Gemma

Multimodal & multiple model sizes

Multimodal models from Google DeepMind, available in a range of sizes.

Release & licence
Featured release
Gemma 4
Licence
Apache 2.0

Gemma 4 uses Apache 2.0. Earlier Gemma releases have different terms.

Read the original release (opens in a new tab)
06

Moonshot AI

Kimi

Coding & tool-oriented workflows

A mixture-of-experts model built around coding and tool-oriented work.

Release & licence
Featured release
Kimi K2 Instruct
Licence
Modified MIT

Modified MIT is not standard MIT. Read the release-specific conditions.

Read the original release (opens in a new tab)
13

MiniMax

MiniMax

Coding, search & tool use

Open-weight models for coding, research and multi-step work with tools.

Release & licence
Featured release
MiniMax-M2.5
Licence
Modified MIT

The MiniMax model licence modifies MIT. Read the terms for this specific release.

Read the original release (opens in a new tab)
14

NVIDIA

Nemotron

Efficient reasoning & adaptable thinking

Hybrid reasoning models with adjustable thinking budgets and support for retrieval workflows.

Release & licence
Featured release
Nemotron Nano 9B v2
Licence
NVIDIA Open Model Licence

This release uses the NVIDIA Open Model License Agreement, not Apache 2.0 or MIT.

Read the original release (opens in a new tab)
15

Baidu

ERNIE

English, Chinese & multimodal research

An open-weight family spanning language and vision, with English and Chinese language models.

Release & licence
Featured release
ERNIE 4.5 21B A3B PT
Licence
Apache 2.0
Read the original release (opens in a new tab)
16

Xiaomi

MiMo

Efficient reasoning, code & long context

Xiaomi's open-weight models combine efficient attention with reasoning, coding and tool-oriented workflows.

Release & licence
Featured release
MiMo-V2-Flash
Licence
MIT
Read the original release (opens in a new tab)
17

Swiss AI Initiative / Switzerland

Apertus

Multilingual AI & transparent research

Swiss-developed language models sharing weights, training data and recipes for transparent multilingual research.

Release & licence
Featured release
Apertus 8B Instruct 2509
Licence
Apache 2.0

Developed by ETH Zurich, EPFL and CSCS. The named September 2025 release uses Apache 2.0.

Read the original release (opens in a new tab)
18

EuroLLM consortium / Europe

EuroLLM

European languages & translation

European-built multilingual models covering all 24 official EU languages, with a focus on translation and instruction following.

Release & licence
Featured release
EuroLLM 9B Instruct
Licence
Apache 2.0

Developed by a European research consortium. Hugging Face requires accepting the download conditions for this release.

Read the original release (opens in a new tab)
19

Tencent

Hunyuan

Adaptive reasoning & long-context work

Tencent's open-weight models combine fast and slow thinking modes with long-context understanding and tool use.

Release & licence
Featured release
Hunyuan-A13B-Instruct
Licence
Tencent Hunyuan Community Licence

This custom licence excludes the EU, UK and South Korea. Inclusion here is not approval for deployment.

Read the original release (opens in a new tab)
20

Meituan

LongCat

Dynamic compute & efficient tool use

Meituan's mixture-of-experts models use dynamic expert activation for efficient text generation and tool-oriented tasks.

Release & licence
Featured release
LongCat-Flash-Chat
Licence
MIT
Read the original release (opens in a new tab)
21

Aleph Alpha / Germany

Kolibri

German/English reasoning & tool use

German-developed open-weight reasoning models focused on German and English, long-context work and tool calling.

Release & licence
Featured release
Kolibri-1
Licence
Apache 2.0
Read the original release (opens in a new tab)

Model names and trademarks belong to their respective owners. Our support for the open-source and open-weight movement is not an endorsement of, partnership with, or approval of the companies behind these models, their conduct or their actions, especially where they conflict with Underleau's values. Inclusion does not imply approval for any particular deployment. Underleau is in development; production model selection and testing are still ahead.

03 / CAPABILITY, NOT A PADLOCK

Open does not mean less capable.

There is no rule that a model must be closed to be good. Published results show open-weight models competing with proprietary systems. The honest comparison is always a specific model, task and test.

A concrete example.

In DeepSeek's original R1 evaluation, R1 scored slightly higher than OpenAI o1-1217 on two mathematics tests, while o1 scored higher on GPQA Diamond. Open weights did not prevent competitive performance; neither model won everything.

These historical results illustrate what is possible. They are not a current leaderboard, an independent Underleau evaluation or a guarantee for your documents.

Original evaluation and methodology (opens in a new tab)
DeepSeek-reported results, January 2025. Pass@1 (%); higher is better.
BenchmarkDeepSeek-R1Open weightsOpenAI o1-1217Closed weights
AIME 202479.879.2
MATH-50097.396.4
GPQA Diamond71.575.7

Author-reported settings: up to 32,768 generated tokens; 64 samples to estimate pass@1, temperature 0.6, top-p 0.95. Benchmark conditions and task coverage matter.

04 / OPEN ABOUT THE DETAILS

Different kinds of open.

We use the words carefully. Celebrating a community does not mean overlooking the differences.

Open weights

The trained parameters are available under a stated licence. That does not automatically include training data, the complete training process or unrestricted commercial use.

Open-source AI

A broader commitment to the materials and permissions needed to study, modify and reproduce a system. Projects such as Ai2's Olmo share substantially more than a finished checkpoint.

Responsible deployment

Neither label guarantees accuracy, privacy or safety. Hosting, provider policies, access controls and application-level safeguards still matter. Your data is not made public just because model weights are public.

OUR COMMITMENT

Built on shared progress.
Held to a high standard.

We want Underleau to build on commercially usable open-weight models. We will choose for real-world quality, source-grounded answers, latency, privacy and safety, not just the lowest token price. Retrieval, tools and verification matter as much as the model behind them.

We support the researchers and builders who share their work: giving them credit, linking to original releases and building with openness in mind.

Build the conversation with us