내가 좋아할 만한 콘텐츠
Building a foundation model is an art! it involves many complex dimensions and stages, from architecture, data, pre-training, post-training to inference. getting it all right requires masterful and tasteful execution. there are very few teams around the world that can make…
Check out this fun work that explores fine-tuning small LFM2 and Qwen3 models for French. arxiv.org/abs/2510.05846
📚 Efficient Language Specialization for Small Language Models @maxencelsb and @SinoueG have released a preprint about their excellent work on fine-tuning small models in French. It shows a solid post-training pipeline to improve French performance while preserving English…
📚 Efficient Language Specialization for Small Language Models @maxencelsb and @SinoueG have released a preprint about their excellent work on fine-tuning small models in French. It shows a solid post-training pipeline to improve French performance while preserving English…
new LFM for PII Japanese data extraction on par with GPT5! 💪🏽 enjoy
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…
LFMs fly on iphone! you can try a large series of LFMs on Apollo: apps.apple.com/us/app/apollo-…
Apple wasn’t kidding, the iPhone 17 Pro is really built for running LLMs Here’s LFM2 8B A1B by @LiquidAI_ running on-device with MLX in @LocallyAIApp, the iPhone runs the 8B model with zero struggle Thanks @Prince_Canuma for the port to MLX, it made the MLX Swift port possible
Day 1 of the @LiquidAI_ fine-tuning hackathon in Tokyo this weekend. Jointly organized with @weights_biases and @LambdaAPI
Liquid AI Releases LFM2-8B-A1B: An On-Device Mixture-of-Experts with 8.3B Params and a 1.5B Active Params per Token How much capability can a sparse 8.3B-parameter MoE with a ~1.5B active path deliver on your phone without blowing latency or memory? Liquid AI has released…
Btw, it should get faster on the next version of MLX Swift. We made some improvements to 1D grouped convs that will speed up this model nicely.
Apple wasn’t kidding, the iPhone 17 Pro is really built for running LLMs Here’s LFM2 8B A1B by @LiquidAI_ running on-device with MLX in @LocallyAIApp, the iPhone runs the 8B model with zero struggle Thanks @Prince_Canuma for the port to MLX, it made the MLX Swift port possible
Hello everyone! Let me (re)introduce myself!
People should give a try on @LiquidAI_ models with Spectrum. You can SFT/RLFT your models with VERY LOW memory footprint, without having to do LoRA or qLoRA... This beautiful thing prevents a lot of the catastrophic forgetting. LFM models work out of the box.
I added LFM 2 8B A1B in @LocallyAIApp for iPhone 17 Pro and iPhone Air The first mixture of experts model by @LiquidAI_, 8B total parameters (1B active), performance similar to 3-4B models but speed of a 1B model Runs great on the 17 Pro with Apple MLX
We just released LFM2-8B-A1B, a small MoE optimized for latency-sensitive applications on-device. Larger model quality with the speed of a 1.5B class model. Huggingface: huggingface.co/LiquidAI/LFM2-… Blog: liquid.ai/blog/lfm2-8b-a…
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…
LFM2-8B-A1B just dropped on @huggingface! 8.3B params with only 1.5B active/token 🚀 > Quality ≈ 3–4B dense, yet faster than Qwen3-1.7B > MoE designed to run on phones/laptops (llama.cpp / vLLM) > Pre-trained on 12T tokens → strong math/code/IF
Small MoEs are on the rise. @LiquidAI_ drops LFM2-8B-A1B.
LFM2-8B-A1B just dropped on @huggingface! 8.3B params with only 1.5B active/token 🚀 > Quality ≈ 3–4B dense, yet faster than Qwen3-1.7B > MoE designed to run on phones/laptops (llama.cpp / vLLM) > Pre-trained on 12T tokens → strong math/code/IF
Enjoy our even better on-device model! 🐘 Running on @amd AI PCs with the fastest inference profile!
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…
Meet LFM2-8B-A1B by @LiquidAI_ - 8B total and 1B active params 🐘 - 5x faster on CPUs and GPUs ⚡️ - Perfect for fast, private, edge 📱/💻/🚗/🤖
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…
LFM2-8B-A1B Liquid AI’s first on-device MoE, with 8.3B total parameters and 1.5B active per token. It matches 3–4B dense model quality while running faster than Qwen3-1.7B. Architecture - 18 gated short-conv blocks, 6 GQA blocks (LFM2 backbone) - Sparse MoE feed-forward layers…
LFM2-Audio-1.5B Liquid AI’s first end-to-end audio foundation model, built for real-time conversation at only 1.5B parameters. Competitive with much larger models, it unifies speech and text without separate ASR or TTS. Architecture - LFM2 multimodal backbone - FastConformer…
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…
United States 트렌드
- 1. Happy Birthday Charlie 25.2K posts
- 2. Good Tuesday 23.4K posts
- 3. #tuesdayvibe 2,494 posts
- 4. #Worlds2025 23.9K posts
- 5. Bears 92.1K posts
- 6. Pentagon 68.5K posts
- 7. Romans 10.5K posts
- 8. Blinken 25.4K posts
- 9. Snell 25.7K posts
- 10. Jake Moody 14.6K posts
- 11. Jayden 23.8K posts
- 12. Trump's Gaza 138K posts
- 13. Commanders 53.7K posts
- 14. Dissidia 4,421 posts
- 15. Joji 35.8K posts
- 16. Conquered 29.8K posts
- 17. Swift 292K posts
- 18. #BearDown 2,493 posts
- 19. Martin Sheen 5,642 posts
- 20. Time Magazine 14.4K posts
내가 좋아할 만한 콘텐츠
-
Symmetry and Geometry in Neural Representations
@neur_reps -
Danijar Hafner
@danijarh -
Jascha Sohl-Dickstein
@jaschasd -
Yi Ma
@YiMaTweets -
Ofir Nachum
@ofirnachum -
ICLR 2026
@iclr_conf -
Jonathan Frankle
@jefrankle -
Taco Cohen
@TacoCohen -
Cohere Labs
@Cohere_Labs -
Karol Hausman
@hausman_k -
Miles Cranmer
@MilesCranmer -
Nathan Lambert
@natolambert -
Paul Liang
@pliang279 -
Johannes Brandstetter
@jo_brandstetter -
Behnam Neyshabur
@bneyshabur
Something went wrong.
Something went wrong.