BottleCap AI: The Prague Lab Teaching Qwen to Think Less, Now on Your Mac

BottleCap AI, the Prague startup co-founded by word2vec author Tomáš Mikolov, has released a Mac-native build of its newest model. On 24 September 2026 it published ThinkingCap-Qwen3.8-27B in Apple's MLX format. It weighs 21 GiB, against roughly 52 GB for the full-precision weights, and fits on a 32 GB Apple Silicon Mac with image input still working.

The company is small, but its pitch is unusual: instead of training ever-bigger models, it makes existing ones cheaper to run. This guide covers who is behind BottleCap AI and what ThinkingCap changes about Alibaba's Qwen. It also shows how to run the model on a Mac, the licence line to read before using it for work, and where the company sits in the Czech AI scene.

The Key Takeaways

• The Mac build: ThinkingCap-Qwen3.8-27B in MLX format is 21 GiB and runs on a 32 GB Mac, with vision and speculative decoding intact.
• What it does: BottleCap AI reports 37.2% fewer thinking tokens than stock Qwen3.8-27B for an average accuracy drop of 0.86 points across 12 benchmarks.
• Who made it: a Prague team led by word2vec author Tomáš Mikolov and Beat Saber co-founder Jaroslav Beck, backed by a $7.5m seed round.
• The catch: the licence is PolyForm Small Business. Personal use and firms under 100 people are fine; larger companies need a commercial deal.
• Its home turf: Czech AI runs on applied companies, university labs in Prague and Brno, and a new €40m national AI Factory.

What Is BottleCap AI?

BottleCap AI is a Prague company founded in 2025 that fine-tunes and compresses language models so they spend less computing power on each answer. It has three founders: Tomáš Mikolov as chief scientific officer, Jaroslav Beck as CEO, and researcher David Herel. Unlike most Czech AI companies, it works on the models themselves rather than on products built around them.

The Scientist: Tomáš Mikolov

Mikolov's story starts in Brno. He finished his PhD on neural-network language models at Brno University of Technology in 2012, then joined Google Brain, where he led the work that became word2vec. The follow-up word2vec paper won the NeurIPS 2023 Test of Time Award, with NeurIPS noting it had been cited over 40,000 times. He moved to Facebook AI Research in 2014 and came home in 2020 to lead a research group at CIIRC, the Czech Technical University's research institute for informatics and robotics in Prague.

The Founder Who Sold Beat Saber

Beck is the unexpected name. He co-founded Beat Games, the studio behind the VR hit Beat Saber, which Meta bought in 2019. According to the startup news site Startup Kitchen, he put $10 million of his own money into BottleCap AI at launch. In January 2026 it closed a $7.5m seed round led by the British fund 20VC. According to CzechCrunch, the round also drew the Czech fund Rockaway Ventures, founders of Canva, ElevenLabs, Hugging Face, Datadog and Supercell, and Ivo Lukačovič, the founder of Seznam, the Czech search engine.

The Idea: Efficiency, Not Scale

From day one the founders talked about training language models up to 100 times more efficiently, and Startup Kitchen reported that the team was not trying to match the tech giants on computing power. In practice that means taking open models people already run and cutting the waste out of them. The company also has its own base model, CAP1, which powers an iOS news app called Pulse, and it credits a EuroHPC grant for time on Italy's Leonardo supercomputer.

What ThinkingCap Actually Changes

Reasoning models like Qwen think out loud before they answer, and they often think far longer than the answer needs. BottleCap AI's ThinkingCap series targets that one habit. It fine-tunes an open model so it reaches the same answer with fewer "thinking" tokens, and deliberately leaves knowledge, answer style and safety behaviour alone.

The base here is Qwen3.8-27B, the compact dense sibling of Alibaba's Qwen 3.8 flagship. According to Alibaba's model card, it has 27 billion parameters, reads images and video natively, and ships under the permissive Apache 2.0 licence. BottleCap AI published the tuned version on 22 September, with these headline numbers at the highest reasoning setting:

Source: BottleCap AI's release post. Every figure is the company's own, run on one Nvidia H200 with identical settings for both models.

Where the Savings Cost You

Math is the price. AIME 2026 drops almost four points, the largest loss in the table. Long-context retrieval went the other way and improved by 2.25 points, outside the base model's margin of error.

The fairer comparison is Qwen's own dial. Qwen3.8-27B already lets you lower its reasoning effort, but in BottleCap AI's tests dropping stock Qwen to "medium" cut thinking by 52.1% and cost 9.16 points of accuracy. ThinkingCap gets most of that saving for under one point. On a Mac, where every token is generated by your own chip, fewer thinking tokens means you wait less for the same answer.

How to Run ThinkingCap on a 32 GB Mac

This is the part that put the model in front of Mac owners. BottleCap AI announced it on X on 24 September:

ThinkingCap: Qwen3.8-27B now runs on a Mac and fits a 32GB machine!

4-bit MLX build - 21 GB, down from 52 GB
Performance is very close to the full precision version.
Vision capabilities and MTP speculative decoding are kept.

https://huggingface.co/bottlecapai/ThinkingCap-Qwen3.8-27B-NVFP4

— BottleCapAI (@BottleCapAI) September 24, 2026

One correction before you download: the link in that post goes to the NVFP4 build, which is made for Nvidia data-centre GPUs and will not run on a Mac. The Mac version is a separate repository, bottlecapai/ThinkingCap-Qwen3.8-27B-MLX-4bit-DWQ.

Which Build Is Which

The MLX build is not a plain 4-bit squeeze. The large MLP blocks are 4-bit, but attention, the last eight layers' MLPs and several sensitive projections stay at 8-bit, and the vision tower stays at full precision. BottleCap AI fitted the quantisation to the full model's own outputs, a method called DWQ. On its one-seed tests the Mac build stays within about half a point of full precision on GPQA-Diamond, 87.4% against 87.9%.

The Three Commands

You need Python on an M-series Mac with 32 GB of unified memory or more. Install the vision-capable MLX runner first. The plain mlx-lm package also loads the model, but BottleCap AI warns it drops image input and speculative decoding:

pip install mlx-vlm

Then run a prompt. The first run downloads the 21 GiB of weights:

python -m mlx_vlm generate \
  --model bottlecapai/ThinkingCap-Qwen3.8-27B-MLX-4bit-DWQ \
  --prompt "Explain the Cauchy–Schwarz inequality." \
  --enable-thinking --max-tokens 4096

For extra speed, switch on MTP speculative decoding. The model's own next-token head drafts ahead, so there is no second model to download:

python -m mlx_vlm generate \
  --model bottlecapai/ThinkingCap-Qwen3.8-27B-MLX-4bit-DWQ \
  --draft-model bottlecapai/ThinkingCap-Qwen3.8-27B-MLX-4bit-DWQ/mtp-drafter \
  --draft-kind mtp --draft-block-size 3 \
  --prompt "..." --enable-thinking --max-tokens 4096

BottleCap AI measures 2.41 accepted tokens per round on the Mac build, almost identical to the full model's 2.43. On an M5 Mac, the oMLX runner has an optional INT8 prefill mode, and oMLX measured it reading a 32K-token prompt 34% faster on an M5 Max. It is off by default and can change outputs.

Keep the memory maths in mind. The weights alone take 21 of your 32 GB, and macOS and the model's working memory need the rest, so close heavy apps before a long session. Shopping for a machine to do this? Our guide to the best MacBook for AI covers the memory tiers. Our look at what the M5 Ultra Mac Studio runs locally covers the top of the range.

Or Skip the Download

Running a model locally is the right call when privacy or offline use matters. It is a poor trade when you just want a good answer, because a 27-billion-parameter model on a laptop is still smaller than the flagship models in the cloud. Fello AI puts Qwen next to Claude, GPT, Gemini, DeepSeek and Grok in one native Mac app, so you can send the same question to each and keep your 32 GB free. It also runs on iPhone and iPad, with a free tier and a paid plan at $9.99 a month. On the Mac there is also a native Qwen desktop client built on it.

The Licence Line to Read First

ThinkingCap is open to download, but it is not open source in the Apache sense its Qwen base is. BottleCap AI releases it under the PolyForm Small Business License 1.0.0, with an extra grant for personal, non-commercial use by individuals.

The small-business test is concrete. Under the PolyForm terms, your company qualifies if it has fewer than 100 employees and contractors combined and less than $1 million in revenue in the prior tax year. That dollar line is set in 2019 money and adjusted for inflation. Above either line, commercial use needs a separate licence from BottleCap AI. The upstream Qwen parts stay under Apache 2.0. If the difference between open weights and open source is new to you, our explainer on what open source AI actually means covers it.

The model card also carries an access form asking for your name, company and email. It doubles as a sales channel: the card pitches enterprise versions that go further, and BottleCap AI's release post offers fine-tunes for lower reasoning settings on a customer's own workload.

Where BottleCap AI Fits in Czech AI

Search for Czech AI and you get a national strategy from 2019, a few non-profit hubs and a list of software agencies. The real scene is small and research-heavy, and BottleCap AI is its outlier: most of the country's best-known AI companies make AI useful inside a business rather than working on the models.

A Lineage That Starts With a Word

The Czech Republic has been in this story from the start. The word "robot" comes from Karel Čapek's 1920 play R.U.R. Čapek first wanted the Latin labori, and took "roboti" at the suggestion of his brother Josef, whom he later named as the word's real inventor.

The Companies Worth Knowing

The Czech names that travel are mostly applied businesses. They sell AI into a workflow rather than selling a chatbot:

Funding figures come from E2B's own announcement, SiliconANGLE and EU-Startups.

E2B is the one a developer on a Mac is most likely to have touched without knowing it. It was founded by Czech engineers and now operates from San Francisco, and its homepage lists Claude Code and Codex among the agent setups that run on its sandboxes.

The Public Infrastructure Behind It

The state side got real money in 2026. The Czech AI Factory launched in Ostrava on 12 May 2026. According to the consortium's launch release, its budget is nearly CZK 1 billion (about €40m), half from the EU's EuroHPC programme and half from the Czech government. Roughly half of that pays for the KarolAIna supercomputer. One of its first projects is a voicebot for the 112 emergency line, now in testing with the Fire Rescue Service.

Language is the other front. Charles University's Jan Hajič leads OpenEuroLLM, a consortium building open models for 32 European languages that Expats.cz reported had secured €34m. It sits in the same sovereignty conversation as Mistral, Europe's best-known AI lab, but with a university consortium in place of a startup.

The Verdict: BottleCap AI Wins by Making Models Cheaper

The Czech Republic will not out-spend the US or China on training runs, and BottleCap AI is not trying to. Its bet is that the next useful gain comes from removing waste. ThinkingCap is a clean first proof: a third less thinking for under a point of accuracy, measured carefully and published with every number. The Mac build turns that from a data-centre story into something you can test on your own desk.

If you have a 32 GB Apple Silicon Mac and a personal project, download it and try it on the kind of questions where stock Qwen rambles. If you work at a company above the licence line, treat it as a demo and talk to BottleCap AI before building on it. For the rest of the open field, our ranking of the best open-source AI models shows where a 27B model sits.

FAQ