Minish (@minishlab) profile picture

Building Model2Vec, Semble, SemHash, and the Potion embedding models. GitHub: https://xitter.gq/t.co/WAoFTUEQ6O. HuggingFace: https://xitter.gq/t.co/1khDD6Y4YB

Joined January 2025
Photo by Minish (@minishlab)Photo by Minish (@minishlab)Photo by Minish (@minishlab)Photo by Minish (@minishlab)Photo by Minish (@minishlab)Photo by Minish (@minishlab)Photo by Minish (@minishlab)Photo by Minish (@minishlab)Photo by Minish (@minishlab)Photo by Minish (@minishlab)
Minish retweeted
⚡ @minishlab's Blink static models bring fast CPU embeddings in 8M (English) and 64M (100+ languages) sizes! Both give you 256d vectors and work directly with Sentence Transformers. Here's blink-8M finding the right help article 🧵
1
5
52
1,283
I have a confession to make: I am addicted to finding new ways of making your search pipeline faster and more performant. And we all know the main pain point of any search pipeline: switching embedding models. When you have millions of points, your model starts to be too slow on cheap CPUs, so you often need to switch to expensive GPUs💰 But there's a way out: it's called static embeddings, and, for my latest blog post at @qdrant_engine in collaboration with @minishlab, we managed to embed 32.9M MIRACL passages in 35-40 minutes with their potion models running in Fastembed (which we heavily optimized for static models). Blog link👇
1
4
23
777
Our second blink release in a week! Today we're releasing blink-multilingual-64M 🚀 Last week's blink-8M was English-only. blink-multilingual-64M brings the same recipe to more than 100 languages, and it's the successor to potion-multilingual-128M at half the size. 🧵
1
2
1
10
675
Features: - Highest MMTEB mean of all static models we compared at 256 dimensions - Big jumps over potion-multilingual-128M: retrieval goes from 41.22 to 47.52, STS from 61.15 to 67.47 and bitext mining from 40.72 to 57.03 (!) huggingface.co/minishlab/bli…
3
66
How small can we go with the hardware we run our embedding models on? Turns out, pretty small: in my latest blog post at @qdrant_engine, I took potion-base-2M by @minishlab and, with some no_std Rust 🦀, I made it run on a ESP32 microcontroller with 16MB of memory. And it ran pretty fast as well! It embedded a dozen tokens in approximately 1.2ms, meaning a throughput of 830 embeddings/s (assuming very short texts). Of course, this is far from the 11-15k embeddings/s that I saw static models produce on my Mac with 48GB of unified memory (more on that very soon 👀) , but it's pretty impressive considering it uses 3000x less memory. Also, it's proof that we can start thinking about running search 𝘵𝘳𝘶𝘭𝘺 on the edge. Link to the blog below👇
3
8
2
16
1,466
Have you ever generated 32.8 million embeddings in 35 minutes? We did, using the potion static embedding models from @minishlab, that we recently integrated in Fastembed (our framework for local embeddings inference). In our latest article in collaboration with Minish, we dive into what static embeddings are, what optimizations we introduced in Fastembed to make them go brr, and what tradeoffs they carry in terms of retrieval quality. Link below👇
3
6
23
923
We've just released our newest model: blink-8M 🚀 It's a drop-in replacement for potion-base-8M, and it beats every static model we compared on MTEB at 256 dimensions, including potion-base-32M and potion-retrieval-32M. 🧵
3
1
3
23
1,451
Features: - 7.7M parameters, 31MB on disk - Highest MTEB mean of all static models we compared at 256d (55.35, up from 53.30 for potion-base-8M) - SOTA on multiple tasks. Retrieval goes from 31.22 to 38.86 (!) compared to potion-base-8M huggingface.co/minishlab/bli…
3
117
We’ve just released Semble v0.3.0! Biggest changes: - Automatic disk caching: First search builds the index; later searches are (much) faster. - Search more than code: new `--content` flag for searching docs, config, code, or all of them together. 🧵
1
1
4
175
- JSON output: CLI and MCP results are now structured JSON, making Semble easier to use from agents and scripts. - Better file control: use `.sembleignore` to exclude files or force-include custom extensions. Release notes: github.com/MinishLab/semble/…
1
95
Today we're releasing Semble, a fast and accurate code search library built for agents 🤖! We're also releasing potion-code-16M, a small code-specialized static embedding model that powers Semble. 🧵
1
1
2
11
883
Main features: - Fast: indexes a full codebase in ~250 ms and answers queries in ~1.5 ms, all on CPU - Accurate: on par with transformers - MCP server: drop-in tool for Claude Code, Cursor, Codex, OpenCode, and any other MCP-compatible agent - Zero setup: no API keys, no GPU 🧵
2
1
156
Model2Vec 0.7.0 is out now, as well as a blogpost on model size reduction techniques! This release features a large number of ways to improve the distillation process. - Vocabulary quantization - Configurable pooling - A number of small improvements and bugfixes 🧵
2
1
1
8
443
We also have a new blogpost on model size reduction where we showcase how to reduce model size by a factor of 15, creating a 6MB model (!) without impacting performance. Links: Release notes: github.com/MinishLab/model2v… Blogpost: minish.ai/blog/2025-10-05-si…
1
133
semhash is so so so convenient and fast
2
1
2
16
2,031
We have a new website (and name): minish.ai We’ve been working on an improved website for a while, and it’s finally here. It has documentation for all our packages as well as our blog. More things coming soon! 🚀
1
9
268
Some guy forked our "model2vec-rs" crate, and put it under the "model2vec" name on crates io and then didn't tell us about it. See here: crates.io/crates/model2vec Like what's the goal here except name squatting.
1
1
5
300
We just released model2vec 0.6.0! This is a big release, containing many big improvements 🔥 GitHub release: github.com/MinishLab/model2v… PyPi release: pypi.org/project/model2vec/0… 🧵
1
1
3
142
- Model improvements: nearly all distilled models will perform better, especially in STS and clustering tasks. A while ago we published a blog post on modernbert not working, but we now found out why, and fixed it! 🧵
1
1
56
- Smaller tokenizers: all tokenizers are now 40% smaller, at no cost to anyone. A blog post with experimental results is coming in the next couple of days.
53