BottleCap AI

No GPU needed. ThinkingCap Qwen3.6-27B (FP8) is now available to everyone through a simple API on Requesty. Get building in 3 steps: 1) Sign up: app.requesty.ai/sign-up 2) Search for sference/thinkingcap-qwen3.6-27b 3) Build your application Why through Requesty: - OpenAI-compatible API: integrate in 3 lines of code, works with the SDK you already use - Production-ready from day one: 99.99% uptime, automatic failover, EU data residency - Full cost control: real-time analytics, caching, and spending limits built in Launch pricing: $0.40 per million input tokens, $3.00 per million output tokens. Prefer to run it yourself? Weights are free on Hugging Face. 🤗 https://lnkd.in/dGPt-J9S Give it a try and tell us what you build!

We are so happy about the widespread adoption of our first ThinkingCap model, with the official GGUF version reaching over 300k downloads on Hugging Face. We are working hard to satisfy the demand for more quantized versions, releasing an official FP8 quant with more to follow soon. In the meantime, a rapidly growing pool of community contributions is ready for you to explore. The greatest strength of our token-cutting speedup is that it works in perfect synergy with existing inference optimization techniques. Combining ThinkingCap with FP8 quantization and Qwen's native multi-token-prediction led to a 6.25x average speedup over vanilla Qwen 3.6 inference in our limited internal testing. We continue exploring how far we can push this and are preparing a comparison with other popular Qwen 3.6 forks, so stay tuned! 🔗ThinkingCap: https://lnkd.in/epVua-8X 🔗GGUF: https://lnkd.in/dqzRRtvk 🔗FP8 Quant: https://lnkd.in/dGPt-J9S 🔗Community Contributions: https://lnkd.in/dANeQBkD

Blown away. Our ThinkingCap model is trending on 🤗Hugging Face with 300,000 downloads in just 2 days. And that’s for the quantized version. Glad you like it! 🙌 We’re also planning to release more quantized versions soon. https://lnkd.in/dqzRRtvk

Is the tokenmaxing era over? More tokens do not always mean better results. The future of AI will not be about wasting as much tokens as possible, but more about getting best possible outcome for the best possible price. That is why we released ThinkingCap: a model series focused on cutting unnecessary reasoning for specific tasks while preserving quality. - Same answer, faster. - Fewer tokens burned. - More value from the context window you already have. 🔗 Full technical write-up with methodology & numbers in our blog post: https://lnkd.in/eXTbjP28 🔗 HuggingFace: https://lnkd.in/epVua-8X #AI #LLM #Efficiency #ThinkingCap #BottleCapAI

🚨 Introducing efficiency focused “ThinkingCap” AI model series.  Starting with a 2× thinking token reduction on average in Qwen 3.6 27B, with up to 10x faster generation on individual examples. It's live on HuggingFace! Our core thesis at BottleCap AI has always been: 𝗘𝗳𝗳𝗶𝗰𝗶𝗲𝗻𝗰𝘆 𝗶𝘀 𝗻𝗼𝘁 𝗮 𝘁𝗿𝗮𝗱𝗲𝗼𝗳𝗳, 𝗶𝘁'𝘀 𝗮𝗻 𝗮𝗱𝘃𝗮𝗻𝘁𝗮𝗴𝗲. Faster answers, lower cost, better throughput without sacrificing capability. We finetuned Qwen3.6-27B, to reduce the thinking tokens without losing accuracy of the output. As a side effect, the model answers appear to also be shorter and more to the point, without degrading the quality. And this is just the beginning. The results across twelve out-of-domain benchmarks: 🏆 𝗨𝗽 𝘁𝗼 𝟭𝟬𝘅 𝗳𝗮𝘀𝘁𝗲𝗿 in specific responses 📉 ~𝟰𝟲% 𝗳𝗲𝘄𝗲𝗿 𝗿𝗲𝗮𝘀𝗼𝗻𝗶𝗻𝗴 𝘁𝗼𝗸𝗲𝗻𝘀 on average 🎯 −𝟬.𝟳𝗽𝗽 𝗮𝗰𝗰𝘂𝗿𝗮𝗰𝘆 𝗱𝗶𝗳𝗳𝗲𝗿𝗲𝗻𝗰𝗲 (comparable accuracy) 🔄 𝗙𝗲𝘄𝗲𝗿 𝗿𝗲𝗮𝘀𝗼𝗻𝗶𝗻𝗴 𝗹𝗼𝗼𝗽𝘀 & 𝗱𝗲𝗮𝗱-𝗲𝗻𝗱𝘀 ⚡ 𝗟𝗼𝘄𝗲𝗿 𝗹𝗮𝘁𝗲𝗻𝗰𝘆, 𝗹𝗼𝘄𝗲𝗿 𝗶𝗻𝗳𝗲𝗿𝗲𝗻𝗰𝗲 𝗰𝗼𝘀𝘁, better real-world throughput We're releasing the checkpoint publicly on HuggingFace under Apache 2.0. This is our first ThinkingCap model release and more is coming as our algorithms can be applied to most open models. 🔗 Try it yourself: https://lnkd.in/epVua-8X 🔗 Full technical write-up with methodology & numbers in our blog post: https://lnkd.in/eXTbjP28 #AI #Efficiency #OpenSource #LLM #BottleCapAI #ThinkingCAP

Great to finally meet F1 champion Nico Rosberg 🏎️ Nico is an investor in BottleCap AI, so it was cool to catch up in person. Even more special for me: this was my very first time seeing and hearing a live F1 race in person. So. Damn. Loud!! Thanks for an invite 🙌

The Czech National Bank successfully tested BottleCap AI’s tool for efficient AI adoption in line with EU regulatory requirements. A meaningful step for us, and one we're proud of.

Come say hi next week 👋 two of our ML engineers are out in the wild, sharing findings straight from the lab. Vojtěch Bartek is at Google Cloud Summit in Prague with a lightning talk: Not Every LLM Is Someone You'd Want on Your Team. It's about how post-training and alignment can quietly shift how a model behaves, and how to test, compare, and govern LLMs before you trust one with a real decision. Good one to catch if you care how these models actually behave once they're in your stack. 🗓 𝟬𝟵 𝗝𝘂𝗻𝗲, 𝟰:𝟰𝟬 𝗽𝗺 · 𝗣𝗿𝗮𝗴𝘂𝗲 𝗖𝗼𝗻𝗴𝗿𝗲𝘀𝘀 𝗖𝗲𝗻𝘁𝗿𝗲 Register here → https://lnkd.in/dz5K9DHP Karol Lasocki joins the Mews R&D panel: What Happens When AI Leaves the Lab. The messy, honest gap between "it works in the demo" and "it works in production." 🗓 𝟭𝟬 𝗝𝘂𝗻𝗲, 𝟲 𝗽𝗺 · 𝗠𝗲𝘄𝘀 𝗛𝗤, 𝗣𝗿𝗮𝗴𝘂𝗲 Register here → https://lnkd.in/dUBs2_sT If you're around for either, come find us.

BottleCap AI made TechCrunch’s Top 21 Startups to Watch in Europe 🚀 From day one, our goal has been simple: invest in real foundational research, turn what we learn into products we genuinely want to use ourselves and keep tightening the loop between research and the real world. In the last 3 months, we shipped two products: • Pulse Community News: built on top of our first foundational model "CAP1" giving people deeper context on the topics they follow so they’re less likely to be misled by clickbait headlines. • AI Scan: focused on measuring how models have been modified after their original internet-scale training across politics, culture, finance, and more. This is still just the beginning, but it matters to us that we’re building from our own work, not just wrapping existing models. That’s why research is at the core of BottleCap. Europe has to decide what role it wants to play in AI. If we keep building mostly wrappers, we stay downstream from the foundational technology. If we want to matter long term, we need to build innovation from model architecture and efficiency to new AI products and the infrastructure around them. It’s time to build. Thanks TechCrunch & Julien Codorniou for including us 🙌 https://lnkd.in/dMd46YHe

Quoting TechCrunch. Not a bad way to start the week. Happy to be featured in their list of 21 European startups to watch, alongside teams we deeply respect. Thanks to Julien Codorniou (20VC) for backing us, and to Anna Heim for the feature. See the full list → [https://lnkd.in/eWuzhijx]