Ternlight: 7MB Embedding Model Runs in Your Browser
Ternlight packs a full embedding model into 7MB and runs it entirely client-side via WASM. Here's how the browser-based AI tool works and why it matters.
> **TL;DR:** Ternlight is a 7MB embedding model that runs entirely inside a web browser via WebAssembly, letting developers generate embeddings and run semantic search without sending any data to a server. It gained 175 points on Hacker News as developers highlighted it as a privacy-preserving, offline-capable alternative to server-hosted embedding APIs.
Key Takeaways
- Ternlight compresses an embedding model down to roughly 7MB, small enough to load and run in a browser tab. - It executes fully client-side via WebAssembly (WASM) — no API calls, no server round-trip, no data leaving the device. - The demo drew 175 points on Hacker News, with developers framing it as a case study in lightweight, privacy-preserving AI tooling. - Because embeddings are computed locally, Ternlight can power semantic search and similarity matching even when the browser is offline. - It arrives alongside a broader industry push toward smaller, cheaper, more efficient models rather than ever-larger ones.
What Is Ternlight?
Ternlight is an embedding model small enough to ship inside a web page. At around 7MB, it's compact enough to download alongside a typical modern website's assets, load in the browser, and run without ever contacting a backend server. A live demo of the model is available at [ternlight-demo.vercel.app](https://ternlight-demo.vercel.app/), where it converts text into vector embeddings directly in the client.
Embeddings are the numerical representations that power semantic search, recommendation systems, and retrieval-augmented generation (RAG) — they let software compare the *meaning* of two pieces of text rather than just matching keywords. Normally, generating embeddings means calling a hosted model over an API. Ternlight's pitch is that this step can happen entirely on-device.
How a 7MB Model Runs Entirely Client-Side
Ternlight uses WebAssembly (WASM) to execute inside the browser's own sandbox, rather than shipping requests to a remote inference server. WASM lets code written in lower-level languages run at near-native speed in any modern browser, which is what makes it feasible to load and execute a machine learning model directly on a user's machine — whether that's a laptop or a phone — without a dedicated GPU server behind it.
The practical effect: a developer can drop Ternlight into a web app and get semantic search or text-similarity features working the moment the page loads, with no network dependency once the model itself has downloaded. That also means the text being embedded never leaves the user's device, which is a meaningfully different privacy posture than sending queries to a third-party embeddings API.
Why Developers Are Paying Attention
Ternlight picked up 175 points on [Hacker News](https://news.ycombinator.com/item?id=48811644), a strong showing that signals real interest from the developer community rather than passing curiosity. The appeal breaks down into a few concrete advantages:
- **Privacy by default** — since inference runs locally, no user text is transmitted to a server for embedding. - **Offline capability** — semantic search and matching keep working without an internet connection once the model is loaded. - **No inference costs** — there's no per-call API bill, since the computation happens on the client's own hardware. - **Simple deployment** — a small model file is easier to bundle and cache than standing up a hosted inference endpoint.
For smaller apps, browser extensions, or privacy-sensitive tools, that combination removes a whole category of backend infrastructure most embedding-based features currently require.
Ternlight vs. Server-Hosted Embedding Models
Most production embedding models today are hosted behind an API and run on server-side hardware, often because of their size and compute requirements. Ternlight's approach inverts that: by aggressively shrinking the model, it trades some of the raw retrieval quality larger, server-hosted models can offer for the ability to run anywhere a browser does — with zero server costs and zero data leaving the device. That's a real tradeoff, not a free lunch, but for many lightweight use cases — in-page search, local document matching, offline note-taking apps — it may be a tradeoff worth making.
Part of a Wider Efficiency Push
Ternlight's compact footprint fits a pattern showing up across the AI industry this year: a growing push toward doing more with less compute. Debates over inference costs and margins have intensified since [GLM 5.2 sparked an 'AI margin collapse' discussion on Hacker News](https://speka.info/blog/glm-5-2-sparks-ai-margin-collapse-debate-on-hn), while frontier labs continue racing on both capability and efficiency — from the rumored [GPT-5.6 Sol Ultra for OpenAI Codex](https://speka.info/blog/gpt-5-6-sol-ultra-rumored-for-openai-codex) to Anthropic's recent [Claude Sonnet 5 launch alongside the relaunched Fable 5](https://speka.info/blog/claude-sonnet-5-ships-as-anthropic-relaunches-fable-5). Ternlight sits at the opposite end of that spectrum from frontier-scale models: rather than pushing capability upward, it pushes footprint downward, aiming for a model small enough to disappear into a browser tab entirely.
What to Watch Next
Ternlight is currently a demo project, not a shipped product with an established roadmap, so it's worth watching whether the model gets packaged as a reusable library, how its retrieval accuracy holds up against larger hosted embedding models in real-world tests, and whether other developers build on the same on-device approach. For more coverage of emerging developer tools like this, check speka.info's [New AI Tools & Skills hub](https://speka.info/new-ai-tools/).
Frequently Asked Questions
What is Ternlight?
Ternlight is a compact, roughly 7MB embedding model that runs entirely inside a web browser using WebAssembly, allowing developers to generate text embeddings and perform semantic search without a server.
How does Ternlight run in the browser?
It uses WebAssembly (WASM) to execute directly on the client device, so embeddings are computed locally instead of being sent to a remote API for processing.
Why does a browser-based embedding model matter?
Because it removes the need for server-side inference: there's no data transmitted off the device, no per-request API costs, and semantic search features can keep working offline.
Is Ternlight as accurate as large hosted embedding models?
That tradeoff isn't fully established yet — a model this small likely sacrifices some retrieval quality compared to larger, server-hosted embedding models in exchange for its tiny footprint and on-device execution.
Where can I try Ternlight?
A live demo is available at ternlight-demo.vercel.app, and the discussion around it can be found in its Hacker News thread.
Sources
- https://ternlight-demo.vercel.app/ - https://news.ycombinator.com/item?id=48811644
