Bonsai 27B runs a 27B model on your phone — and Anthropic launches Claude for Teachers
Prism ML's Bonsai 27B squeezes a 27-billion-parameter model onto a phone. Anthropic also shipped Claude for Teachers, a purpose-built education product.
Prism ML's Bonsai 27B squeezes a 27-billion-parameter model onto a phone. Anthropic also shipped Claude for Teachers, a purpose-built education product.
Prism ML just released Bonsai 27B, a 27-billion-parameter model that the team claims runs on a phone. It hit #2 and #7 on Hugging Face's trending models list and pulled 642 points on Hacker News — making it the most cross-confirmed story today. If the on-device claim holds up, this is a big deal for builders who want local inference without a GPU.
Prism ML published two GGUF variants on Hugging Face: Bonsai-27B-gguf and Ternary-Bonsai-27B-gguf. The ternary version uses 1.58-bit weights, which is how they get the file size small enough to fit on mobile hardware. Both are available to download now. The announcement page at prismml.com/news/bonsai-27b has the technical details. CBW has not run either model locally yet, so the phone-inference claim is unverified on our end.
Anthropic launched Claude for Teachers, a product aimed at K-12 and higher-ed educators. The announcement is live at anthropic.com/news/claude-for-teachers. Anthropic did not publish pricing details in the launch post. If you build tools for schools or sell to education buyers, this signals Anthropic is moving into that vertical directly — which is both an opportunity (they're validating the market) and a competitive signal.
OpenAI posted a guide titled 'How to manage AI investments in the agentic era' aimed at enterprise buyers. It's a positioning document more than a product launch, but it tells you where OpenAI thinks the enterprise conversation is heading: ROI measurement for autonomous agents, not just chat. Worth a skim if you're pitching AI tools to business clients.
Google DeepMind CEO Demis Hassabis posted on X about a plan to develop AI safely. The post got 150 points on Hacker News. No new product or policy was announced — it's a statement of intent. We're noting it because Hassabis rarely posts publicly about safety strategy, so the signal is worth watching.
Three infrastructure tools shipped updates this week. Crawl4AI hit v0.9.2 — it's a web-scraping library built for feeding AI pipelines. LanceDB released python-v0.35.0-beta.2, a beta of their vector database Python client. Weaviate shipped v1.38.4, a patch release of the open-source vector database. None of these are headline features, but if you're running a RAG pipeline, check the changelogs.
A paper called 'Read It Back' (arXiv 2607.11886) argues that pretrained multimodal language models can act as zero-shot reward models for text-to-image generation — meaning you don't need a separately trained reward model to score image quality. Early result, but relevant if you're building image generation pipelines and want a cheap quality filter.
Separately, the Bank for International Settlements published a working paper on how AI companies are financing the current boom — shifting from cash flows to debt. It's dry reading but useful context if you're watching AI company burn rates or thinking about where the infrastructure investment cycle is headed.
1. Download Bonsai-27B-gguf from Hugging Face and run it with llama.cpp on your laptop or phone. Test whether the ternary variant actually fits in your device's RAM and compare output quality to a standard quantized 7B model you already use.
2. If you sell to schools or tutoring platforms, read the Claude for Teachers launch post and map which features Anthropic is offering directly. Find the gap — what do teachers need that a general product won't cover — and build that.
3. Upgrade Crawl4AI to v0.9.2 and use it to scrape a public dataset (job listings, product reviews, news headlines) into a LanceDB vector store. You'll have a working local RAG prototype by the weekend.
Confirmed: Bonsai 27B and Ternary-Bonsai-27B GGUF files are live and downloadable on Hugging Face as of today.
Not independently verified by CBW: We have not run Bonsai 27B on a phone or any mobile device. The on-device inference claim comes from Prism ML's own announcement page.
Confirmed: Anthropic published the Claude for Teachers announcement at anthropic.com/news/claude-for-teachers. Pricing was not disclosed in the launch post.
Worth noting: The OpenAI agentic-era investment post is a thought-leadership document, not a product launch. No new features or pricing were announced.
Not independently verified by CBW: The BIS paper on AI financing is a working paper, not a peer-reviewed study. Treat the debt-cycle framing as one analyst's view, not settled fact.
Source: Prism ML — Bonsai 27B announcement — https://prismml.com/news/bonsai-27b
Source: Hugging Face — Bonsai-27B-gguf — https://huggingface.co/prism-ml/Bonsai-27B-gguf
Source: Anthropic — Claude for Teachers — https://www.anthropic.com/news/claude-for-teachers
Source: OpenAI — Managing AI investments in the agentic era — https://openai.com/index/managing-ai-investments-in-agentic-era
Source: Hugging Face Papers — Read It Back (arXiv 2607.11886) — https://huggingface.co/papers/2607.11886
Source: BIS — Financing the AI boom (PDF) — https://www.bis.org/publ/bisbull120.pdf
Get tomorrow's best AI project in your email. With a guide that works. Free. No spam.