Edoardo Mosca

ML Research Scientist @LiquidAI
Working on Liquid Foundation Models (LFMs): post-training

[Edoardo Mosca @EdoardMosca]

LFMs running at 20ms end-to-end are going places!

[Liquid AI @liquidai]

We're celebrating an exciting milestone in our partnership with @Shopify : our LFMs have now processed 1 billion requests on Shopify’s platform!

Read more about our multi-year partnership here: https://t.co/t8mH7AcPxq

[Liquid AI @liquidai]

Introducing LFM2.5-230M: our smallest model yet, built to run fast anywhere (CPUs, NPUs, and GPUs) to enable agentic tasks on phones, robots, home and network automation devices.

> 230M parameters, built on the LFM2 architecture
> Pre-trained on 19T tokens, with a 32K context extension
> Post-trained with distillation from LFM2.5-350M
> 213 tok/s decode speed on Galaxy S25 Ultra (CPU)
> 42 tok/s on a Raspberry Pi 5 (CPU)
> Competes with and often beats models more than twice its size on instruction following, data extraction, and tool use.
> use it for large-scale data extraction pipelines or lightweight on-device agentic workloads.

🧵

[Liquid AI @liquidai]

Introducing LFM2.5-Embedding-350M and LFM2.5-ColBERT-350M: two multilingual retrieval models built for ultra-fast and accurate search across 11 languages.

> End-to-end retrieval latency as low as 1.5ms with our enterprise stack! 🚀

> Consistently best-in-class multilingual and cross-lingual performance across Arabic, German, English, Spanish, French, Italian, Japanese, Korean, Norwegian, Portuguese, and Swedish.

🧵

[Edoardo Mosca @EdoardMosca]

@kadirnardev @paulabartabajo_ Assuming that's per device, and looking at your screenshot -> you are training on 162,307,968 samples. 400hours (16,6 days) seems fairly reasonable for an MoE on a single node.

[Edoardo Mosca @EdoardMosca]

@kadirnardev @paulabartabajo_ hey are you using TRL? what batch size?

[Liquid AI @liquidai]

The LFM2 Tech Report is now live on arXiv!

We share everything from our novel hardware-in-the-loop architecture design, pre-training, and knowledge distillation, to the post-training recipe for small models.

> 🤗LFM2 class of models has over 3.3M downloads
> ⚛️LFM2 nanos from 350M to 8.3B MoE
> 👁️Vision-language capabilities (LFM2-VL)
> 👄👂Multimodal speech processing (LFM2-Audio)
> 🗒️Information retrieval (LFM2-ColBERT)

We hope this serves as a useful resource and inspiration for anyone building open and efficient foundation models. 🚀

[Mikhail Parakhin @MParakhin]

After months embedded with @LiquidAI, we're deploying their LFMs in production - the architecture is legitimately different. Sub-20ms inference on real workloads. ~50% fewer parameters, outperforms alternatives, 2-10× faster. No quality compromise.

[Pau Labarta Bajo @paulabartabajo_]

Imagine you can ask the engineers building frontier AI at Liquid AI anything you want.

Well... you can stop imagining.
It's happening in 3 days ⬇️

[Liquid AI @liquidai]

LFM2-8B-A1B has greater knowledge capacity than competitive models and is trained to provide quality inference across a variety of capabilities. Including:

> Knowledge
> Instruction following
> Mathematics
> Language translation

2/n

[Liquid AI @liquidai]

Today, we release LEAP, our new developer platform for building with on-device AI — and Apollo, a lightweight iOS application for vibe checking small language models directly on your phone.

With LEAP and Apollo, AI isn’t tied to the cloud anymore. Run it locally when you want, for speed, privacy, and reliability, using LEAP’s end-to-end toolkit for on-device AI.

1/

[Liquid AI @liquidai]

Introducing LFM-7B, our new best-in-class language model in English, Arabic, and Japanese optimized to be the substrate for private enterprise chat, code, fast instruction following, and agentic workflows. 1/

[Liquid AI @liquidai]

We raised a $250M Series A led by @AMD Ventures to scale Liquid Foundation Models and accelerate their deployment on-device and at enterprises https://t.co/u37Cv9DVa4

[Alexander Amini @xanamini]

Had a wonderful time launching https://t.co/GD6wZ8Heyk alongside the entire @liquidai team!

⭐️⭐️ WATCH the full 3-hour event online today 👉 https://t.co/o4nn75hZlB

We unveiled a lot:
1⃣ #LiquidFoundationModels (LFMs): a new generation of #AI models that achieve best in class quality + efficiency

2⃣ 1B, 3B, and 40B Language LFMs: to deploy intelligence at all scales (from edge-to-enterprise)

3⃣ A suite of #multimodal LFMs: to unlock new AI applications across industries (incl. bio 🧬, driving 🚘, finance 💰, and time-series 📈)

4⃣ Edge LFMs: for offline + private environments. we demo running LFMs entirely on a phone, with *no* internet connection, 100% private

5⃣ Flexible speech interfaces: end-to-end speech models for fast interfaces to our models (from chat to structured json outputs)

6⃣ Special fireside talks + partnerships: @MassGovernor, @SebastienBubeck, @MParakhin, @AMD, @Samsung, @ArenaBioworks, @Deloitte, @Capgemini, @CTC_Press

Great team effort from @ramin_m_h, @mlech26l, @jimmysmith1919, @maximelabonne and entire @liquidai team!

[Mikhail Parakhin @MParakhin]

Just to reiterate: https://t.co/QLUZL0KcR7 model is the first one I've seen that managed to break away from the prediction made by @ilyasut in 2020. Literally everyone else is in the epsilon vicinity of Ilya's graph below.

[Liquid AI @liquidai]

Today we introduce Liquid Foundation Models (LFMs) to the world with the first series of our Language LFMs: A 1B, 3B, and a 40B model. (/n)

[Ramin @ramin_m_h]

today, I want to share the core values that shape our culture at Liquid. here we go:

no-bullshit meritocracy,
burn the playbook,
proactive execution and purposeful ownership,
be white-box explainable and
let's grow together.

Allow me to elaborate:

-------------
A CULTURE OF EXCELLENCE AT LIQUID

as we continue our journey to create very capable AI that solves real problems at every scale, I've been reflecting on what makes Liquid unique.

This mission is ambitious; it requires an exceptional team operating at the highest level. That's why I want to reaffirm and clarify the core values that define our culture:

1. “No-bullshit” meritocracy

anyone who wants to stay long at Liquid should be working on something that is on the critical path. at Liquid, results speak louder than anything else. we value ideas based on their merit, not their source.

2. Burn the playbook

AI is a new frontier - there is no playbook. tear down the status quo, innovate, and rebuild from first principles. never do anything just because “that’s the way it’s done.” be comfortable with extreme ambiguity.

3. Proactive execution and purposeful ownership

because everyone at Liquid is an expert in some domain, there is a high level of trust given to employees to execute autonomously within their domain and deliver something that works, the first time. if something isn’t working step outside your comfort zone and fix it, otherwise delegate, defer, and don’t interfere.

4. Be white-box explainable

We build white-box models within a white-box organization. at any point in time, every employee at the company knows their relevant inputs (what they
consume from others), their outputs (what they produce for others), and, when collaborating, can explain exactly what they’re doing and why.

5. We grow together

employees at Liquid prioritize the needs of the company, and the company prioritizes the needs and well-being of its employees. Liquid is a product-
driven, customer-first business and its problems are everyone’s problems: company goals should align with and support the personal, professional, and academic desires of individuals on the team.

as we move forward, focusing on making AI solutions more accessible and integrating them efficiently across enterprises, these core values will be our north star.

I'm proud of the culture we've built at Liquid, and I'm excited to see how it will continue to evolve and strengthen as we grow. together, we're not just building AI – we're shaping the future of problem-solving at every scale.

read more about the culture at @liquidai below, and if these values resonate with you consider joining us.

https://t.co/kyPZsMOXej

Ramin

[Edoardo Mosca @EdoardMosca]

Detecting Word-Level Adversarial Text Attacks via SHapley Additive exPlanations
https://t.co/XHg52hgizq

GrammarSHAP: An Efficient Model-Agnostic and Structure-Aware NLP Explainer
https://t.co/T28GkJ1x3W

(3/3)

[Edoardo Mosca @EdoardMosca]

It's been a terrific week at #ACL2022 ! I am extremely grateful for the amazing #NLP talks, the beautiful city of Dublin, and most importantly for all the incredible people I met.

Once more thanks to my coauthors - without them this wouldn't have been possible. (1/3)

[Edoardo Mosca @EdoardMosca]

Don't forget to check out our three papers!

“That Is a Suspicious Reaction!”: Interpreting Logits Variation to Detect NLP Adversarial Attacks
https://t.co/QHk7bAqoCU
(2/3)

[Thomas Wolf @Thom_Wolf]

ML researchers: we don’t really need to learn about biology or developmental psychology because plane don’t fly like birds anyway

Also ML researchers: “this new algorithm take inspiration form [our layman view of] how XXX happens in humans”

🙃