--- title: AIpster emoji: ๐Ÿง  colorFrom: indigo colorTo: purple sdk: static pinned: false --- # AIpster **An independent think tank on artificial intelligence, society, and the future of thought.** We're a collective of computer science friends from the late '90s who turned a WhatsApp group into a laboratory for exploring what AI is doing to how we work, build, and think. ๐ŸŒ [aipster.com](https://aipster.com) --- ## What we do here This Hugging Face organization is where we publish the **artifacts** of our exploration โ€” models, datasets, and tools that come out of the experiments we write about on our blog. We're not a company. We don't sell anything. We build things to understand them, then share what we learned. --- ## Focus areas - ๐Ÿ”ฌ **Small specialist models** โ€” distillation, fine-tuning, and the art of making tiny models punch above their weight - ๐Ÿงญ **Prompt engineering & routing** โ€” how prompts become infrastructure, not just text - ๐Ÿ› ๏ธ **Local LLM workflows** โ€” what 96 GB of VRAM can (and can't) do - ๐Ÿค– **Coding agents & automation** โ€” how AI is reshaping software development from the inside out - ๐Ÿ“– **AI & society** โ€” the uncomfortable conversations the industry would rather skip --- ## What you'll find here ### Models **[DevRouter-1.5B](https://huggingface.co/aipster/DevRouter-1.5B)** โ€” our first release. A tiny prompt router that reads a raw developer prompt and returns a single JSON decision: a cleaned-up rewrite, an `intent` / `complexity` classification, a suggested model-tier `route`, and the context the prompt forgot to include. Built on Qwen2.5-Coder-1.5B (Apache 2.0) and distilled from a stronger teacher, it holds **~96% valid-JSON** and runs at **~280 tokens/s on a single RTX 3090** โ€” small enough to sit in front of your real models and triage every prompt in 1โ€“3 seconds. - ๐Ÿง  [aipster/DevRouter-1.5B](https://huggingface.co/aipster/DevRouter-1.5B) โ€” fp16 weights (transformers / vLLM) - ๐Ÿ“ฆ [aipster/DevRouter-1.5B-GGUF](https://huggingface.co/aipster/DevRouter-1.5B-GGUF) โ€” Q8_0 + F16, plug-n-play with Ollama / llama.cpp And one honest caveat, because we ship those too: **Q6 and below quantizations break its JSON.** A small model doing strict structured output is far more fragile than the "Q4 is fine" rule of thumb suggests โ€” ship Q8_0 or F16. ### Datasets *Coming soon* โ€” curated and synthetic datasets from our distillation experiments, released alongside the models that use them. ### Spaces *Coming soon* โ€” interactive demos of our experiments. --- ## Read our work - ๐Ÿ“ [Blog](https://aipster.com) - ๐Ÿงช [How we built our distillation pipeline](https://aipster.com) *(coming soon)* - ๐Ÿ” [Four GPUs, Two Weeks, and the Uncomfortable Truth About Local LLMs](https://aipster.com/four-gpus-two-weeks-and-the-uncomfortable-truth-about-local-llms/) - ๐Ÿค– [I Stopped Learning n8n. I Just Told My Coding Agent What I Wanted](https://aipster.com/i-stopped-learning-n8n-i-just-told-my-coding-agent-what-i-wanted/) - ๐Ÿ’ธ [Two Hours to Mass Extinction: What Coding Agents Mean for the Open-Core Business Model](https://aipster.com/two-hours-to-mass-extinction-what-coding-agents-mean-for-the-open-core-business-model/) --- ## Philosophy > We build to understand. We share to learn together. Everything we publish here is open. Code, weights, datasets, methodology โ€” including the failures. Especially the failures. --- ## Get in touch - ๐ŸŒ Website: [aipster.com](https://aipster.com) - ๐Ÿ“ฌ Email: contact@aipster.com --- *Independent. Curious. Slightly skeptical.*