I wrote an article summarizing how we designed our tokenizer back in early 2024:

It's a good tokenizer, sir

We recently ported LFM2 by @LiquidAI_ to Cactus (YC S25), the 350m-i8 runs at 188 tokens/sec on M4 CPU-ONLY. Gemma3 270m-i8 runs at 170 tokens/sec for reference. On an old iPhone 13 Pro, it should near 100 tokens/sec, no NPU or GPU! It’s officially one of our recommended models

Meet LFM2-VL-3B, our latest on-device VLM. Top scores in multi-modal instruction following

We have a new nano LFM that is on-par with GPT-5 on data extraction with 350M parameters. Introducing LFM2-350M-PII-Extract-JP 🇯🇵 Extracts personally identifiable information (PII) from Japanese text → returns structured JSON for on-device masking of sensitive data. Delivers

Day 1 of the @LiquidAI_ fine-tuning hackathon in Tokyo this weekend. Jointly organized with @weights_biases and @LambdaAPI

It's a good model sir. Very proud of the team, we worked very hard to be on the Pareto frontier of quality and efficiency. Even had the chance to write a CPU-optimized kernel for MoE to squeeze everything from the hardware, and that gave us those sweet throughput results.

LFM2-8B-A1B: Our MoE model that runs on a phone. This is just the start, much more to come ...

Meet LFM2-8B-A1B, our first on-device Mixture-of-Experts (MoE)! 🐘 > LFM2-8B-A1B is the best on-device MoE in terms of both quality and speed. > Performance of a 3B-4B model class, with up to 5x faster inference profile on CPUs and GPUs. > Quantized variants fit comfortably on

https://t.co/mZqPqGSjeK

We achieved strong LLM performance + blazing fast edge inference with just: - Grouped Query Attention (global sequence mixer) - Double Gated short convolutions (local sequence mixer) - No linear attention/SSMs needed

Just wrote my first Substack post: "Flipping the Script: Why Short Convolutions Don't Need Linear Attention" TL;DR: Everyone's asking why linear attention needs short convolutions. We asked the opposite: do short convolutions need linear attention? LFM2 proves they don't 🎯

Cool release by @LiquidAI_: LFM2-Audio-1.5B It’s a pretty cool omni-architecture that enables prediction of both text and audio tokens, meaning it can handle multi-turn S2S, ASR, and TTS (with voice description) within a single model. Great to see, once again this year, a model

Our @LiquidAI_ LFM2-Audio-1.5 in a nutshell: - both text and audio in - both text and audio out - 1.5B -> runs locally - open-weight license

I have a new blog post about the so-called “tokenizer-free” approach to language modeling and why it’s not tokenizer-free at all. I also talk about why people hate tokenizers so much!

We continue to scale our @LiquidAI_ LFM2 series of super efficient language models with LFM2-2.6B

The secret sauce most definitely is in the data, given that the architecure is fairly standard: Qwen3 backbone + NaViT SigLip2 (i.e. it uses packed vision sequences). They use patch_size=16 and pixel_shuffle_scale_factor=2 in order to use few image tokens. A 256x256 image will

We trained our @LiquidAI_ LFM2-350M model 1400x beyond "compute optimal" > Chinchilla scaling laws: ~20 tokens per param > LFM2-350M: ~28,000 tokens per param (1400x more) Why? Because Chinchilla only concerns training compute, while we care about inference cost

Very proud of our team at Liquid AI Japan! We’ve just released our first Japanese task-specific SLM in the model library ( https://t.co/3CFVqrAF01), with many more to come. It’s a small 350M model (i.e. reaches 200 tok/s prefill 40tok/s decode on a Raspi5), so you may notice a

I wrote a short article about LFM-2's (by @LiquidAI_ ) hybrid architecture w/ illustration + simple pytorch impl.