Fernando Fernandes Neto
@FernandoNetoAi
Machine Learning and AI researcher wayyy before all this hype.
116 Following 1.4K Followers
It is amazing having such high-caliber folks partnering with you to make sure you are the shaping the frontier of SLMs @ArtificialAnlys @liquidai
SO SO beautiful :)
Today, we release LFM2.5-VL-3B, a lightweight vision-language model that reads screens, documents, and the physical world. It handles digital screens across mobile, web, and desktop, grounds objects to coordinates, reads text and charts, and calls tools from either text or image input.
Built on LFM2.5-2.6B base, with a SigLIP2 400M NaFlex vision encoder
> Pre-trained on ~34T tokens
> Vocab size: 128K
Comparable or better scores compared to models up to 2.6x its size:
> ScreenSpot-v2 80.7, ahead of Gemma-4-E4B at 51.2
> RealWorldQA 73.1, ahead of InternVL-3.5-4B at 67.7
> TextVQA 84.3, ahead of Qwen3.5-4B at 81.2
> RefCOCO-avg 87.9, up from 57.1 on LFM2-VL-3B
> ToolSandbox 59.5, up from 26.4 on LFM2-VL-3B
🧵
Show more
LFM2.5-2.6B now runs locally on your Mac 🚀
Great release by @LiquidAI — and Nativ supports it Day 0.
Full BF16, no quantization:
⚡️ 11,000+ tok/s prefill
⚡️ 82 tok/s decode
🧠 Under 8.5GB peak memory — even at full 128K context
📈 Scales to 476 tok/s aggregate across 16 concurrent requests
Try it on Nativ 👉
Github repo 📷
Show more
Today we are releasing LFM2.5-2.6B, a very strong on-device agentic model.
We continued pre-training to 34T tokens, expanded the vocab (as in and utilized a 4-stage post-training pipeline to unlock the strong agentic capabilities for its size.
Show more
Congrats to @liquidai on LFM2.5-2.6B! Excited to have partnered with them for Day 0 support in Nativ 🎉
Built for agentic + coding workflows — and it’s fast.
On an M5 Max (48GB) with Nativ v0.2.2 — full bf16, no quantization:
⚡ 11,231 tok/s prefill
⚡ ~84 tok/s decode
🧠 Full 128K context in just 8.5GB
📈 476 tok/s aggregate decode at batch 16
Local inference doesn't get much better than this.
Get started today👇🏽
Show more
LFM2.5-2.6B is OUT!
I LOVED post-training this model, especially working on agentic RL. An exciting part was training inside real agent harnesses like OpenClaw and Hermes. <1.7GB Q4 runs on a phone.
Incredibly proud of the team, can't wait to see what people build with it 🥳
Show more
When @liquidai really builds for the edge <3
The new LFM 2.5 2.6B model by @liquidai is now available on iPhone and iPad.
A fast model designed for edge devices with comparable or better scores compared to models up to nearly 4× its size.
Update your app now.
Show more
This was one of the most challenging but also most pleasant model releases to be working on.
It is incredible what we could do with OPD and RL in such a small model. Plugging the harnesses into the RL stage brought us so much performance!!
It is amazing to be able to be working closely with @songdng @maximelabonne @timseyde @SinoueG @nathanrchn @helloiamleonie @mlech26l @ramin_m_h
Show more
@liquidai AI just released LFM2.5 encoders small bidirectional models with 8k context.
I've added a training recipe for fine-tuning them on @huggingface Jobs (works for any encoder), and used it to train a classifier that suggests task categories for Hub datasets from their READMEs.
One command, 20 minutes on an A100, about $0.80.
Show more
One thing I like about @opencode is the variety of models on the Go plan, but deciding which one to use every time gets annoying.
So I spent some time experimenting with @liquidai LFM2.5-Encoder-350M-Prompt-Router model.
I know routing between LLMs isn't really the model's intended use case, but its zero-shot routing head made for a fun experiment. A single encoder pass (~250ms) classifies the prompt and routes it to one of the 12 models available in OpenCode Go plan.
Still early, but it's been a fun way to get more out of the Go plan.
Show more
New set of LFM2.5-Encoders just dropped.
These are super fast, even at long context and on CPU.
Our team adapted them from their LFM2.5 backbones:
1. Replaced the causal attention mask with a bidirectional one
2. Made the LFM short convolutions non-causal
3. Trained with a masked language modeling objective
The result is a set of two tiny, super fast encoders.
Encoding a document of 12 to 15 pages (about 8k tokens) takes less than 30s on a CPU.
Release blog:
Models on Hugging Face:
Show more
We just released two new encoder models (MLM) in 2026 🚀🚀🚀
They're super fast, easy to train, and strongly multilingual.
Try them today: we created 5 demos on @huggingface
In strong support of open-weight AI and American AI leadership, we are proud contributors to the open-source community and excited to announce that Liquid Foundation Models (LFMs) have surpassed 40 million downloads by the community!
Going forward, we remain committed to accelerating the open-weight release of the next generation of lightweight, powerful LFMs to the world. excited to see what you build with them!
Show more
The economics of local models only make sense in two scenarios:
1. local deployment on consumer / edge device
2. when sovereignty / control is a must
It doesn't make any sense deploying locally any one of these huge open source models if you don't fit any of these 2 criteria
1/n
Show more
