Hobby project, I wanted to "ship a useful model in a web browser". so I distilled a small sentence encoder from MiniLM with ternary quantization-aware training. Also wrote the inference engine from scratch and shipped in Rust → WASM SIMD.
It's an embeddings model, not an LLM: text goes in, a 384-dim vector comes out, and cosine similarity between two vectors tells you how related the texts are — regardless of shared words ("reset my password" ↔ "I forgot my password" → 0.88). Used for semantic search, FAQ/intent matching, and clustering. Running it on-device means search-as-you-type semantic search is performant with no API dependencies.
Nice, I'm really interested in using this for simple semantic search in a native desktop application.
Any comparisons with other tiny embedding models? Did you start from MiniLM-L6 because it's an especially good model in its class? It's hard to figure this out since all you provide is "Retrieval (SciFact NDCG@10)".
But the claimed performance seems way off, I get only 35 emb/sec in firefox on a i5-4570 rather than 400/sec. Is there an issue with falling back to a non-SIMD path? I'll try a native Rust binary next.
Same!
I’m trying to find small models that can embed effectively to enable BM25/hybrid search over a large number of documents for a personal information repository. Ideally, it should run on consumer hardware.
bge-small-en-v1.5 is one that is comparable and what we’re working with for now.
0.84 Spearman fidelity to the MiniLM teacher at ternary precision is a striking result. How much of that is the quantization-aware training doing the work, versus what a post-training ternary quant of the same encoder would give you?
It's entirely the QAT. The whole distillation process is quantization-aware from the start, so the ternary weights are learned rather than fitted after the fact.
The only post-training quantization I applied was int4 on the embedding layer, and I ran a small ablation there to find the sweet spot between size and quality.
gte-small outscores all-MiniLM-L6 on MTEB (~61 vs ~56 avg per the GTE paper). MiniLM is ternlight's teacher (ternlight holds 0.84 Spearman fidelity to teacher).
I haven't run a head-to-head yet; STS-B/MTEB numbers are on the roadmap. Also on the roadmap is to distill gte-small as teacher.
but also maybe you could put a button on the landing page to trigger the demo because it's a bit startling to hear my fans go crazy when opening a webpage.
I really love the coil whine of my GPU (a 5090 FE) when it’s doing LLM stuff. I can hear the different stages, like prefill and decode, and the sounds actually make me reminisce about dial up.
Cool project!
I tried something similar a while ago [1] - I wanted to load up an embedding model and semantically order texts, all in the browser.
So I pull ONNX weights from HuggingFace (MPNet, MiniLM), use Transformers.js to embed, and use a clusterer from scikit-learn (running on pyiodide - it was a surprise to me that this worked flawlessly) on the page - all client-side.
Thank you for this! Local models will bring privacy at some point, and I already know an excellent use case for such a small embedding model (cheap and fast search in a product base). Relying on the CPU is also a plus in my case.
This is a known issue and I am actively trying to find why this is happening. So far it's pretty good on Brave/Chrome.
Tested on Macbook Pro M1 8gb RAM and Macbook Air M1 8gb RAM. Mostly likely because of M series of chips. All tests were done on Brave/Chrome.
Does not work on iPhone 11 Pro Max and iPhone 16 Pro. Mostly likely because of A series of chips. Tests were done on Safari and Chrome and it crashes on both.
If you think about it, running a crypto miner without being asked is probably less annoying than downloading an entire LLM, but only the first will get you in jail.
Interesting project. Happy to see someone who shares an interest in tiny vector embeddings models. I've worked on tiny (1MB - 4MB, 250K - 950K parameters) embeddings models called BERT Hash https://huggingface.co/blog/NeuML/bert-hash-embeddings
This doesn't add any malware risks beyond what a JavaScript-enabled browser already allows.
Re excessive browser memory use: Yes, it adds non-negligible weight, but again, you could already achieve excessive browser memory usage before this. For comparison, a true color 1080p image, uncompressed (which is needed for actual display on screen) is only slightly smaller at 6.22Mb.
That's... how the web works? You download things on demand.
There are JS files larger than 7MB in the wild. They run on JIT engines that displayed severe CVEs over the years. PDFs, video running directly on special hardware encoders. That's the web now.
Very cool! I'd love to point it at my own corpus to index/embed. Would be cool if you could give it a link to a markdown file or even a website to crawl.
It's an embeddings model, not an LLM: text goes in, a 384-dim vector comes out, and cosine similarity between two vectors tells you how related the texts are — regardless of shared words ("reset my password" ↔ "I forgot my password" → 0.88). Used for semantic search, FAQ/intent matching, and clustering. Running it on-device means search-as-you-type semantic search is performant with no API dependencies.
Demo (2k React docs, fully on-device): https://ternlight-demo.vercel.app
Two tiers on npm: - @ternlight/base (7 MB, ~5 ms/embed, more capable embedings) - @ternlight/mini (5 MB wire, ~2.5 ms/embed).
Bundled for Node and browsers.
Repo - see technical details (MIT, training pipeline included): https://github.com/soycaporal/ternlight
Curious if this is something useful, what are the use cases for on-device embeddings.
reply