Richie Suarez

AI Engineer @ Liquid AI

[Richie Suarez @liquid_richie]

@halr9000 set up paperless-ngx on his server, bought a duplex ADF scanner, and uses LFM3-VL-2B for OCR and metadata extraction on CPU.

Document intake is a good fit for a small VL model. https://t.co/k6q53q6fxo

[halr9000 @halr9000]

Feeling overwhelmed by paper, so I bought a duplex ADF scanner & setup paperless-ngx on my server. 1-click and done! LFM3-VL2B is powering OCR and metadata extraction; so small it runs in CPU! Next up: #Hermes workflow integration. Great model @liquidai !

[Richie Suarez @liquid_richie]

@_JoshuaTimothy's Agentic Binning runs LFM3-VL-2B on the Snapdragon NPU via @Qualcomm GenieX: it names, describes and tags each photo and video. Tap Bin Files to sort them into folders.

Analysis is offline after a one-time model download. Code: https://t.co/enWuc8TA0e https://t.co/z7XElLrfFE

[Joshua Timothy @_JoshuaTimothy]

It's official, Agentic Binning for @Snapdragon 📱 is released! Analyze and sort your photos and videos using @liquidai LFM3 on the NPU. This happens entirely on device and offline thanks to @Qualcomm GenieX! Try it out and let me know what you think!
https://t.co/7U6UGQBIJw

[Richie Suarez @liquid_richie]

Transformers now has a shared decision-model layer: pipeline("decision"), model.decide(), DecisionConfig. llama.cpp serves the same /v1/systemone API.

Models keep their own internals. d1's examples use model.system_one(); serving, training and evals can target one contract.

[Lysandre @LysandreJik]

A pretty important new primitive just landed in 🤗 Transformers: **decision models**.

I think this could matter much more than adding support for any single model.

Today, different decision models all approach "making a decision" differently:

- @typesafeai’s Jev, the OG: typed decisions + calibrated probabilities, single forward
- @AutoTrustAI’s JEV builds open System One models around a learned decision block, w/ multimodal.
- @jaredpalmer’s KEV uses a pointer head over the options, sharing the computation of the state across questions.
- @Cloudflare’s CLEF uses a joint schema head that routes evidence between the state, questions and options, scoring the whole schema together.
- @perplexity_ai’s Decider uses a separate 255-way readout on top of Qwen, with non-causal attention in its full-attention layers.
- @liquidai’s d1 is exploring small, multimodal decision models that can run all the way down to the edge.
- And @XunzhuoLiu's Semantic Router team is pushing the serving side hard with System One + the D3 family: one typed interface for choice, score and yes/no decisions, with models from sub-1B to 27B.

That experimentation is good, and needed! But without a common interface, every training framework, inference engine and evaluation tool has to understand every implementation separately.

So Transformers is introducing a shared decision-model layer:

`DecisionConfig`
`pipeline("decision")`
and, most importantly:

`model.decide(...)`

The idea is **not** to make every decision model work the same internally, but for them to speak the same language to the ecosystem around them.

This means that, without running custom or remote code:
- inference engines can serve them through one interface
- training tools can build against common primitives.
- the Hub can package them consistently.
- evaluations can compare them without model-specific glue everywhere.

This is the part of transformers I'm the most excited about: to grow alongside the ecosystem and the many ideas; and to make all of them more usable by everyone.

https://t.co/yslV5Pc8NM

[Richie Suarez @liquid_richie]

Packed room at Bloom's AI forum this week.

Liquid AI Japan demoed an in-car assistant on a 1.5B audio model, Galaxy S24 Ultra CPU, airplane mode: windows, music, a text to the boss, navigation with ETA.

Thanks Patrick H. and Bloom for hosting.
https://t.co/VlPhLSta8h

[Richie Suarez @liquid_richie]

New Liquid AI newsletter is out: LFM3's tri-hybrid architecture, six models including an omni model, many Mercedes-Benz models in North America getting LFMs over the air, and how the team built its own vision encoder to teach a language model to see.

https://t.co/STlmD6pjCV

[Ramin @ramin_m_h]

Today @WSJ covered our work with @MercedesBenz. our flagship LFMs are now rolling out in vehicles in North America, running directly on the hardware inside the car to power in-car assistance! 🌊

this is one of the first deployments of local generative AI in the real world at scale! ✨

seeing our technology become part of someone’s everyday life is one of the most rewarding feelings of building Liquid.

years of work become a small, useful part of someone’s day. they don’t need to know which model is running. they just need it to understand them and work seamlessly.

this is really hard and it took us years of fundamental work from first principles to made happen with a brilliant team of engineers from Mercedes and Liquid.

I believe on-device AI will fundamentally change our relationship with the technology around us. embedded AI is about change of ownership.

With Mercedes-Benz, we are now in production and are building toward deployment at scale. The on-device AI revolution is taking shape in products people actually use, and I’m proud that @liquidai is helping lead it.

To our team and our partners at @MercedesBenzUSA and HQ: thank you for the care, persistence, and trust it takes to make this real.

It feels good to see the work out in the world.

https://t.co/dO11I8dTsv

[Richie Suarez @liquid_richie]

@bebuu_ai PR #31660 names lfm and moves convolution-state snapshots into build_conv_state_store. Its author reports synthetic rollback coverage for lfm. Have you measured decode performance on a Liquid model? https://t.co/8ObCwdWtOn

[Richie Suarez @liquid_richie]

@BehnamEbrahimi Your Tensorguery post compares activation-direction edits with neuron deletion on LFM3-1.7B. Are the scripts public, and how do you measure concept suppression and effects on unrelated behavior?

[Richie Suarez @liquid_richie]

@sharathkrml @fermion_ai @liquidai @GradiumAI LFM3-1.7B uses about 1 GB of memory at 4-bit precision. You report local Phonon STT and TTS around it. How do you handle transcript revisions once reply generation starts? https://t.co/qy24XUd7Ch

[Richie Suarez @liquid_richie]

@marcaruel @liquidai LFM3-Audio-2B supports spoken tool calls and thinking for text-only replies to speech. The README does not document thinking before tool calls. My two-model suggestion changes the single-model setup you want. https://t.co/y22IbXW8VY

[Richie Suarez @liquid_richie]

@MNEMOSYNE_OS @liquidai Your JinjaTemplate/getWeather example reports one call for Paris and reuse on the umbrella follow-up. Does a third turn asking about another city trigger a fresh call with the new city argument?

[Richie Suarez @liquid_richie]

@BennyDaBall_OG Could you share the prompts and harness behind your tool-call slice? I'd add BFCL V4 and IFEval with matched settings to test whether the KLD and BF16-agreement differences carry through to task scores. https://t.co/qy24XUdFrP

[Richie Suarez @liquid_richie]

@termuxSolutions What prompt length did you use for prefill? For a follow-up, try LFM3-1.7B Q4_K_M with llama.cpp v0.8.0+ on the same Vulkan setup and report peak memory plus sustained prefill/decode. The GGUF download is 1.10 GB. https://t.co/5WiFSrS1Y8

[Richie Suarez @liquid_richie]

@MNEMOSYNE_OS @liquidai Your report of working binaries without binding changes is useful for Node users. Could you share a minimal tool-call test alongside the two-turn chat check? Liquid documents LFM3 support in llama.cpp v0.8.0+. https://t.co/5WiFSrRu8A

[Richie Suarez @liquid_richie]

@artistxeature @GillelandKristi LFM3-1.7B was post-trained for Hermes Agent. Try Liquid's Q4_K_M GGUF (1.10 GB download), keep tool descriptions concise, and measure valid calls and task success on your tools before trying lower precision. https://t.co/5WiFSrRu8A

[Richie Suarez @liquid_richie]

@Antonwester5 @xenovacom @liquidai Yes. LFM3-Omni-2B is open weight on Hugging Face, with GGUF, ONNX and MLX builds: https://t.co/aLilzfC5XS

[Richie Suarez @liquid_richie]

@drkpxl_ @MiaAI_lab Florida's rule is 7 p.m. local time, so Central Time counties close at 8 p.m. ET, later, not earlier. Can you share the raw search results Hermes passed to LFM3-1.7B? That shows whether it misread the sources or the data was missing.

[Liquid AI @liquidai]

Adding vision to a language model is more than one extra modality. With large images or long video, the vision encoder can use more compute than the language model itself, so efficient models need a different design. Our VLM team, including @rshubert98, @TianshuYu6, @sina_semnani, and @iamsamstevens, explains how they teach LFMs to see, from mid-training to the custom vision encoder built for LFM3.

[Richie Suarez @liquid_richie]

@AdamZinebi1 Thanks Adam, checking it out!