Post-training @liquidai | PhD @CriteoAILab × @mlia_isir | prev. intern @AIatMeta (FAIR) & @Microsoft
Paris
Joined November 2022
- Tweets272
- Following220
- Followers690
- Likes3.5K
Pinned Tweet
What an amazing time at #EACL2024 last week! 🎉 Thrilled to announce that our paper has been honored with the Best Paper Award! 🏆 Big thanks to the organizers and our team for making this possible! 🙌
🤔 Ever wondered how to summarize an entire book 𝑎𝑙𝑙 𝑎𝑡 𝑜𝑛𝑐𝑒? We delve into this question in our recent paper “𝐋𝐎𝐂𝐎𝐒𝐓: 𝐒𝐭𝐚𝐭𝐞-𝐒𝐩𝐚𝐜𝐞 𝐌𝐨𝐝𝐞𝐥𝐬 𝐟𝐨𝐫 𝐋𝐨𝐧𝐠 𝐃𝐨𝐜𝐮𝐦𝐞𝐧𝐭 𝐀𝐛𝐬𝐭𝐫𝐚𝐜𝐭𝐢𝐯𝐞 𝐒𝐮𝐦𝐦𝐚𝐫𝐢𝐳𝐚𝐭𝐢𝐨𝐧” (#EACL2024). 🧵 (1/8)
LFM3 is finally HERE!!
6 new models spanning text, vision, audio and omni!
Had so much FUN working on agentic RL for our little text model. Still amazed at what we managed to fit into 1.7B parameters.
Super proud of the team and happy to finally share this with the world 🥳
Introducing 𝙻𝙵𝙼𝟹: our new generation of Liquid Foundation Models. With a new tri-hybrid architecture and new multimodal encoders, the LFM3 family is the ideal model suite for specialized intelligence.
> 6 model instances, across text, audio, vision, and omni modalities, generative and decision models.
> 𝙻𝙵𝙼𝟹-𝟷.𝟽𝙱: text, 256k context window.
> 𝙻𝙵𝙼𝟹-𝚅𝙻-𝟸𝙱: vision-language model with video-input support for spatial understanding
> 𝙻𝙵𝙼𝟹-𝙰𝚞𝚍𝚒𝚘-𝟸𝙱: tool calls, voice chat, reasoning in 18 languages.
> 𝙻𝙵𝙼𝟹-𝙾𝚖𝚗𝚒-𝟸𝙱: an omni model combines vision, audio, and language.
> 𝚍𝟷-𝟸𝙱 and 𝚍𝟷-𝙾𝚖𝚗𝚒-𝟸𝙱: decision models with typed outputs and calibrated probabilities in a single pass. Omni adds audio input.
> fine-tune on 1 GPU.
> a new tri-hybrid architecture
> architecture optimized for 14 processor classes across GPUs, CPUs, and NPUs.
> 𝙻𝙵𝙼𝟹-𝚅𝙻-𝟸𝙱 is a beast. It performs competitively with Gemma 4 12B (6x larger), and Gemini 3.5 Flash Lite, and GPT-5.6 Luna, with 1.9B parameters.
> 𝙻𝙵𝙼𝟹-𝟷.𝟽𝙱 is #1 on the Pareto frontier of device intelligence 𝚙𝚒𝚙𝚎𝚝𝚝𝚎.𝚕𝚒𝚚𝚞𝚒𝚍.𝚊𝚒
Every modality for specialized intelligence. Pick the one that fits your use case, customize it, and run it on-device.
1/ 🧵
my boy d1 was at home watching d1-ultra get all the attention 😭
we had to give him a glow-up.
d1 just got an upgrade! Meet 𝚍𝟷.𝟷.
It brings significantly better decision quality than d1, getting closer to d1-ultra while running faster and costing less than half as much.
> built for fast, structured decision-making
> $0.02 / 1M input tokens
> lower latency than d1-ultra
Available today on the Liquid Console: console.liquid.ai
Coming soon to OpenRouter and Vercel.
Having done my PhD at Criteo, I know how challenging it is to run ML inference at scale under strict latency constraints.
Exciting to see what RIPL can unlock for real-time workloads! 🚀
Today, we announce 𝙻𝚒𝚚𝚞𝚒𝚍 𝚁𝙸𝙿𝙻 (real-time inference pipeline), our ultra low-latency inference engine built to put LFMs directly in the request path of real-time workloads like search, ads, fraud, recommendations, and payments.
> complete generated responses end-to-end in sub 50ms
> maintains low latency under continuous peak load
> LFMs get up to 20x faster on RIPL
> RIPL supports Multimodal LFMs
> RIPL supports d1, our decision models too!
RIPL optimizes the full path from request arrival to a complete generated response. As traffic rises, its performance advantage is greatest on the slowest requests. Let’s see some production use cases:
Song retweeted
JevBench v1.7.46 (API offerings only) is live!
Notable changes:
- 🥇 Brand new d1-ultra is leading the pack at Composite Score.
- Lots of strong new model versions in the Top 10.
- Very cheap with good capability: Instinct Dual 4B
- Very fast with good capability: attest-31B-v1.0
Say hi to d1-ultra 👋
Multimodal. 44 languages. Built for the hard decisions.
65 on the @ArtificialAnlys Decision Index. Straight to #1.
SO proud of the team 🥳
Today we release 𝚍𝟷-𝚞𝚕𝚝𝚛𝚊, our most powerful decision model.
> built for hard decision problems.
> d1-ultra is multimodal
> supports 44 languages
> #1 on the @ArtificialAnlys Decision Index (65)
> leads all eight categories of our 112-evaluation suite
1/
LFM2.5-24B-A2B is LIVE!
Absolutely LOVED post-training this model especially pushing its reasoning and agentic capabilities. Really exciting to see what just 2B active parameters can do on laptops and PCs!
SO SO proud of the team, can't wait to see what people build with it 🥳
Today we release 𝙻𝙵𝙼𝟸.𝟻-𝟸𝟺𝙱-𝙰𝟸𝙱, the largest 𝙻𝙵𝙼𝟸.𝟻 family member. It is an efficient local model for laptops and PCs to work through multi-step tasks inside popular agent harnesses.
> 24B total parameters, 2B active per token
> Context expanded to 131K tokens
> A 128K-token vocabulary
> 33T pretraining tokens
> It is a reasoning model
> Very easy to fine-tune and customize
> Performs competitively with Qwen3.6-35B-A3B, a model with ~50% more total and active parameters, on agentic and tool-use tasks.
> Open base and post-trained weights on @huggingface 🤗
1/
Just like LFM2.5-2.6B, LFM2.5-24B-A2B is a pure reasoning models that always thinks before it answers. The <think> tag is prefilled in the chat template. We also released the pre-trained base model.
🤗 Base: huggingface.co/LiquidAI/LFM2…
🤗 Thinking: huggingface.co/LiquidAI/LFM2…
1
11
230
Song retweeted
d1 from @liquidai is live on AI Gateway.
𝚕𝚒𝚚𝚞𝚒𝚍/𝚍𝟷
Make decisions with vision support to classify, route, and score.
vercel.com/changelog/liquid-…
Too easy.
I now challenge @maximelabonne. No pressure 🍿
now available as a HF space 🤗
i challenge @songdng and @SinoueG to beat my score.
(brb getting a new wardrobe)
huggingface.co/spaces/Liquid…
Song retweeted
Local Liquid d1 3B crushed OpenAI Decision at Snake 🐍
@liquidai d1 on a single RTX 3090 decides in 8 ms so its snake made 82 moves a second and ate 86 apples while OpenAI's API took 345 ms a call and ate 3
Run decision models -> atomic.chat
Today we release Open d1: two open-weight multimodal models in our d1 decision model family.
> d1-3B: text + vision
> d1-omni-600M: text + image or text + audio
> Real-time decision making anywhere, from data centers such as @nvidia DGX to RTX workstations to Jetson at the edge.
1/
Can't get enough of d1? Here are two more reasons to be excited:
- d1-3B: text + vision
- d1-omni-600M: text + (image or audio)
running these models locally means going beyond seeing the world to interacting with it in real time (shoutout to @NVIDIARobotics).
wild.
Today we release Open d1: two open-weight multimodal models in our d1 decision model family.
> d1-3B: text + vision
> d1-omni-600M: text + image or text + audio
> Real-time decision making anywhere, from data centers such as @nvidia DGX to RTX workstations to Jetson at the edge.
1/
d1 just got 👀
try it out on console.liquid.ai
coming to Vercel and OpenRouter soon !!
Announcing d1 with vision. 👁️👁️ Our first decision model now supports images, text or both as inputs. We tested d1 against GPT-6.1 Sol and Claude Opus 5.5 on six real applications, from filtering support tickets to inspecting circuit boards. d1 matches or beats GPT-6.1 Sol on four of them. It costs 19x to 200x less than both models and answers significantly faster on every task.
> probabilities for yes/no, choice, or score questions
> one forward pass, without generating tokens
> text decisions in 200 to 300 ms
> Liquid API: console.liquid.ai
🧵
Song retweeted
The same model, with the same weights, scores 62% in one agent harness and 33% in another.
@adithya_s_k and the @huggingface team just released the ultimate guide to multi-harness RL, and it's one of the most practical RL write-ups this year, and everything open!
The trick is simple. Don't touch the harness. Point it at a proxy instead of the model. The proxy speaks all four API formats coding agents use (OpenAI Chat Completions, OpenAI Responses, Anthropic Messages, Gemini). It records the exact token ids and logprobs vLLM sampled, and you train on that. You don't change a single line of Claude Code, Codex or OpenCode.
Results:
🔹 Trained across 4 harnesses at once, LFM2.5-2.6B by @liquidai went from 42% to 54%
🔹 31% fewer tool calls, thanks to a small bonus for solving tasks in fewer steps
🔹 Training in OpenCode alone took OpenCode from 34% to 58%, but the multi-harness model improved everywhere
They also tried the shortcut everyone reaches for: fine-tune on 3,189 successful rollouts from Qwen3.8-27B. Imitation plateaued at 47.5%, below both RL runs. Copying a bigger model doesn't get you there. Practice does.
The best part is that everything is open: the capture proxy in OpenEnv, the trainer in TRL, the tasks, the SFT data, the training code and all seven trained models.
Agents will run in dozens of harnesses.
Now open models can be trained for each of them, by anyone.
Read it here 👇
huggingface.co/spaces/FineEn…
ABSOLUTE banger of a blog, multi-harness RL FTW ! 🔥
1/ Excited to release The ultimate guide to multi-harness RL
The same model behaves differently in every agent harness. So we built an open way to train any model with RL on any task set, inside the harnesses people actually use, like Claude Code, Codex, and OpenCode, without changing a single line of harness or training code.
Trained across four harnesses, LFM2.5-2.6B went from 42% to 54% with 31% fewer tool calls. 🧵
no more tokenmaxxing to pick an answer. d1 is here.
D1 from @liquidai is live on OpenRouter.
It's a competitive decision model: send your app's state and yes/no, choice, or score questions, and get back typed answers with a probability for every option. Zero data retention.
$0.04/M input, $0 output, 65K context
openrouter.ai/liquid/d1
