See how the mainstream AI models are biased

Building efficiency-first foundational AI

Novel AI algorithms & applications
by our in-house research.
Built in Europe.

Explore our products About us

Backed by 20VC & founders of

More about our $7.5M seed round

Scroll

Founders

Tomáš Mikolov: one of the world's most accomplished AI scientists, with 200k+ citations and 20+ years of experience. His most famous inventions include algorithms for scaling neural network language models to large datasets (gradient clipping), neural text generation that essentially all chatbots are based on, the first ever open source implementation of a neural network language model which started the AI revolution (RNNLM), subword tokenization (predecessor of BPE), and vector representation of tokens (word2vec), among many more.

Tomas Mikolov

Co-Founder & CSO

Jaroslav is an entrepreneur, Co-founder of Beat Games, studio behind the biggest VR game to date - Beat Saber (10M+ copies sold), Jaroslav initiated acquisition by Meta in 2019 and continued steering the studio post-exit. After leaving Meta in 2023, he invested in projects driving global positive impact in tech and multiple other fields.

Jaroslav Beck

Co-Founder & CEO

David Herel is Co-founder and AI Researcher at BottleCap AI, where he works on making large language models significantly faster and more efficient. His research focuses on LLM reasoning and model architecture, with publications at leading AI conferences including ICLR and ECAI.

David Herel

Co-Founder & AI Researcher

ThinkingCap does more with less

State-of-the-art language models that think more efficiently, so you get higher performance at a fraction of the cost.

Live tracker →

Our fifth ThinkingCap and the first on a Google model, finetuned from gemma-4-26B-A4B (Google, 2026) to spend 39% fewer thinking tokens on average, with only a 0.15 point accuracy drop.

39%
fewer thinking tokens on average
69%
fewer tokens on everyday tasks
PRO
88%
fewer thinking tokens for your use case

Read model details

Our fourth ThinkingCap and the largest yet, finetuned from Qwen3.8-Flash-Next (Qwen Team, 2026) to spend 39% fewer thinking tokens on average, for about a point of accuracy loss.

39%
fewer thinking tokens on average
51%
fewer tokens on everyday tasks
PRO
76%
fewer thinking tokens for your use case

Read model details

Our third ThinkingCap and the first beyond Qwen, finetuned from Nemotron-3.5-Lightning-30B-A3B (NVIDIA, 2026) to think 23% less on average while keeping comparable benchmark accuracy.

23%
fewer thinking tokens on average
31%
fewer tokens on everyday tasks

Read model details

Our second ThinkingCap, finetuned from Qwen3.8-27B (Qwen Team, 2026) to spend 37% fewer thinking tokens on average, and up to 65% fewer tokens on everyday tasks, for under a point of accuracy loss.

37%
fewer thinking tokens on average
65%
fewer tokens on everyday tasks

Read model details

Our first ThinkingCap, a finetuned Qwen3.6-27B model (Qwen Team, 2026) that spends 46% fewer thinking tokens on average, and up to 68% fewer tokens on everyday tasks, without any significant loss in performance.

46%
fewer thinking tokens on average
68%
fewer tokens on everyday tasks

Read model details

Enterprise

Frontier AI built by the pioneers of the space.

Custom benchmarks, real-time alerts, security audits, and behavior alignment for organizations that depend on AI. We evaluate models against your specific workflows, fine-tune them for your exact use cases, and ensure strict compliance with the EU AI Act and DORA. We turn AI risk management from a manual burden into a system you control.

Learn more

Discover your favorite model’s hidden preferences

AI Scan measures preferences of leading models on sensitive topics and shows where each one leans.

Try AI Scan

Politicians

NegativeNeutralPositive

Donald Trump

Xi Jinping

Vladimir Putin

Joe Biden

Claude Sonnet 4.6DeepSeek V4 ProGPT-5.5

Work alongside the people who built the foundations of modern AI.

At BottleCapAI, our Prague-based team is developing architecture-first foundational models including algorithms for training efficiency, inference efficiency.

Researchers and product engineers work in the same room, moving fast from research breakthroughs to novel AI products which are implemented in fast succession by businesses and consumers.

We’re looking for people who want to build what’s next, and actually ship it.

Open Positions

If you’re a researcher, engineer, or builder passionate about creating the future of AI, join us. We’re hiring!

Natural Language Processing Researcher

Prague, Czech Republic · Full-time · On-site

Machine Learning Engineer

Prague, Czech Republic · Full-time · On-site

ThinkingCap: Gemma-4-26B-A4B — up to 69% and on average 39% less thinking, at unchanged accuracy

This is the fifth model in the ThinkingCap series. We reduced how much Google's gemma-4-26B-A4B-it thinks, without compromising the quality of its answers.

ThinkingCap: Qwen3.8-Flash-Next — up to 51% and on average 39% less thinking, comparable benchmark accuracy

This is the fourth model in the ThinkingCap series. We reduced how much Qwen's Qwen3.8-Flash-Next thinks, without compromising the quality of its answers.

ThinkingCap-NVIDIA-Nemotron-3.5-Lightning-30B-A3B: up to 31% and on average 23% less thinking, comparable benchmark accuracy

This is the third model in the ThinkingCap series. Building on our two earlier releases, we reduced how much NVIDIA's Nemotron-3.5-Lightning-30B-A3B thinks, without compromising the quality of its answers.

Featured in