Introducing LFM3: Every Modality for Specialized Intelligence | Hacker News Reader

Introducing LFM3: Every Modality for Specialized Intelligence

liquid.ai

173 pointsPhilpax67 comments

One thing I have been using LFM 2.6 for is annotating emails, text messages, and other metadata. I keep it all ingesting into one giant database and annotate it all as it comes in on an RTX A4500. The LFM annotations are basically triage, and then slowly annotations trickle into my halo model based on lowest confidence (right now that's DeepSeek 4.1 Flash 1017) for an upgraded annotation. (Agents can then search this database for context).

Excited to roll out LFM 3 to replace it, looks like a big step up. Right now LFM 2.6 gets trustworthy annotations on just over 40% of data points. Just waiting for these to get into LocalAI.

Have you considered fine-tuning on your dataset?

Seeing the fine-tuning guides that are coming out with LFM3, I am now considering it :)

Would be a fun experiment.

I can highly recommend gliner2.5, been using it for tagging news articles, fine tuned on my home 5090 in a few minutes on 500 articles tagged by gpt 6.1, works fantastically. Deployed on hetzner running on a cpu .

gliner-style models are absolutely crazy, I learned about them a couple weeks ago when a co-worker fine-tuned an LFM encoder for promptable classification. as far as I know, you don't need to fine-tune; you can just prompt for classification. My co-worker's demo is here: https://huggingface.co/spaces/LiquidAI/prompt-routing

You're doing something similar to what I've been prototyping for the automatic metadata enrichment part. So far I've been using deterministic older techniques + manual for my personal data stream, but I'm trying to learn what the state of the art is for efficiently annotating data with LLMs and newer ML.

This seems interesting, but if anyone else is also doing a similar thing, I'd appreciate tips on how to approach it. So far, I think I have some small RAG model in a demo or the very small qwen models. I think fine tuning seems interesting too. Some of the decision models applied in layers (like a binary search on a pre-generated corpus) could be fast enough to help too?

idk too many options, i'm sure this is a solved problem lol

It's definitely not a solved problem, and it has been kicking my ass. I don't just want metadata tagging/categorization, I want something that identifies the networks between the information, graduates topics to actual objects, and routinely consolidates/synthesizes narratives that can be easily looked up.

That's very cool. I'm curious if you need a GPU for your use case (email and text triaging). There are some open source alternatives[0] that run/train on CPU, so you can self-host or run locally (and train on your own data for better accuracy), would something like that work for you?

0: http://jeffyclassify.com/

You might also consider the decision models, d1-2B is very strong and good for calibrated classification. Would be happy to share links/resources if you're interested.

Yes, once it's in LocalAI, I intend to try it out.

How do you determine confidence levels? logprobs?

I initially tried annotating everything three times, then judging the annotation with the same model. That failed predictably, so yeah just logprobs right now, judged on a hand-annotated dataset. Mostly correlated - ~80%.

I worked on this release, happy to answer questions/chat about local AI vs cloud frontier models.

In a few words: Why should I care about this?

What makes it cool? What should make me go "oh yeah let's go"?

From what I can see by skimming the document, I would assume "something something specialized use-cases", but what does that mean? And what does it mean assuming I have attention deficit?

Depends on what you care about.

If you want to do local classification/image processing, I think the d1-2B is amazing, it can run on 1GB of unified RAM.

Now that it's released and I have weekends off again, I'm going to play with the VL-2B model and the d1-2B and see if I can prompt it to control a small Lego robot arm like this https://bricksrl.github.io/ProjectPage/

If you like agents, I think the 1.7B is very strong for its size and can be used to automate small stuff like classifying emails, cleaning your downloads folder, etc.

It can also be fine-tuned for specific use cases with SFT/DPO/GRPO. The blog post talks about fine-tuning for wind-turbine classification. If there's things you want to automate, but GPT-6-Luna is not strong enough, and bigger models are too expensive, you might fine-tune one of the LFM3 models and use it locally for repetitive tasks.

No weekends. We need you to get back to work on the next 8B.

Disclosure: I work at Liquid. For local agents, LFM3-1.7B fits in about 1 GB at 4-bit, holds 256K context, and was post-trained for harnesses like Hermes Agent and OpenClaw. On WebArena-Verified it goes from 3.0 with LFM2.5-2.6B to 38.4. You can fine-tune it for your own task on one GPU.

It's a model designed to be tiny and still functional, unlike other larger models that are just scaled down and then run into problems (doom loops, intelligence/reasoning failure, tool call failure). You use it when you want a fine-tuned tiny model processing text/audio/video on multiple platforms. (I don't work for them but this is my takeaway)

I am seeing the following in the reasoning trace:

> However, the instructions say I must use available tools when answering questions.

On "What animal lays the biggest egg?" with no system prompt

I assume this is something baked into the model through lots of RL with tool use?

(I work at Liquid.) Probably, yes. Per the blog, the last 3.5T tokens of pre-training were a mid-training stage with a larger share of agentic traces and tool calls, and SFT covered tool calling and agentic traces too, because the text model is post-trained to live inside agent harnesses. With an empty system prompt it seems to assume it's in one. I'd try a short system prompt that says no tools are available, and if it still does it, a note on the Hugging Face model page would help the team.

Which specific model? and are tools available? I would say that there is lots of RL with tool use, but that the reasoning traces are not always perfect. In general, models think about things that might be wrong, but will recover from this poor line of thinking.

This was using llama-serve on my local which by default gave it current time & unhelpfully "get_info" which is specific to the browser

The full request was

``` {"messages":[{"content":"What animal lays the biggest egg?","role":"user"}],"return_progress":true,"sse_ping_interval":1,"stream":true,"tools":[{"function":{"description":"Returns the current local date and time in ISO 8601 format, with the IANA time zone name","name":"get_datetime","parameters":{"properties":{},"required":[],"type":"object"}},"type":"function"},{"function":{"description":"Get runtime info (OS name), may call when user asks about local files or shell commands","name":"get_info","parameters":{"properties":{},"required":[],"type":"object"}},"type":"function"}],"reasoning_format":"auto","reasoning_control":true,"backend_sampling":false,"timings_per_token":true} ```

Disabling tools removes the bit about mentioning instructions.

Just an interesting detail I noticed.

Do y'all have any penalty/regularization w.r.t. reasoning traces or is it pure SFT -> RL?

interesting, I didn't see it on the playground or on my local llama-server.

SFT -> RL. In our fine-tuning notebooks, we typically say that rollouts that go past the max context length (too much reasoning) get 0 reward, so there is slight pressure for shorter reasoning, but I don't think we have strong regularization. I think the agreed-upon recipe is to first make the model smarter (longer reasoning), then teach it to reason faster/denser.

Congrats for the achievement!

I'm curious if this model would be useful for OCR or at least text documents parsing (PDF, DOCX, etc.), what do you think?

Yes! LFM3-VL-2B is a very strong OCR model for its size, especially given how general a model it is. You can also fine-tune it with a specific output format and it will dramatically improve. We have a Google Colab here: https://colab.research.google.com/github/Liquid4All/cookbook...

Fantastic, I need to try it! Thanks!!

Is interesting how the live mode is restarting the text every 30 sec, wonder how it can impact a thing that may be agreed over several minutes. Say I would like to get action items out of an hour long recording, am I supposed to go straight from the audio with LFM3-Audio-2B? Or is it transcribe first and give the text to LFM3-1.7B? What if I want to fine-tune the action items routine. How many labelled examples do you estimate such task may need?

It all depends on your hardware and throughput requirements. Suppose

1. I were building for my work laptop (M4 Max, 36GB RAM)

2. I had a batched use cases (transcribing + making action items for all meetings in a day)

I would use the LFM3-Audio-2B for ASR (audio -> text transcript) and then LFM3-1.7B for extracting particular action items and adding them to my tool of choice (Linear, Jira, Notes app, etc).

If you can fine-tune (100+ labeled samples, and a GPU), you likely can fine-tune LFM3-Audio-2B for audio -> action items in a single step.

Thank you for an update. Single step does sound exciting, look forward to check it out once I have the full system and data in place. Would LFM3-Audio-2B provide timestamps and speaker labels for each action item? Did you guys train it for any of those?

unfortunately it's not a strong multi-speaker/diarized output audio model. It's a future goal but it doesn't have a timeline. If you have the hardware, I think fine-tuning might be possible.

Is LFM3-Omni-2B worse or roughly equal at vision tasks compared to LFM3-VL-2B?

In other words: Should I select the model that matches my desired capability set or just use Omni?

unfortunately the models are still stronger without extra modalities. So LFM3-1.7B is better at pure text tasks than all other models. the LFM3-VL-2B is better at text/vision tasks than the Omni, the Audio model is a stronger ASR/TTS model than the Omni, etc.

It's a goal of ours to continue unifying training across modalities so that there is truly one model, instead of just this shared pre-trained backbone.

Is it a parameter count problem? Ie why are you not seeing the cross modality bump in performance?

Has the cross modaility bump been conclusively proven?

my intuition is that cross modality improvements do not really show up in practice, and especially not in small models

Did y'all do any particular work that makes these models great for moderation? One of the issues between cloud models vs local is that often people are trying to push the limit of what their hardware can do and it doesn't leave too much headroom for moderation to make the models a little more constrained for kids or schools to use locally, or just for family in general who don't want accidental unhinged results.

In the past I've noticed that some LFM were way better at moderation than much larger models. Not necessarily prejudging what is acceptable, but allowing you to interrogate content with a series of questions to customize the moderation. Other models I tested were far worse.

Decision models are like the next optimization in that direction, but none of the benchmarks for the popular models ever show how good they are at tasks like content moderation. The smaller and more efficient we can get that workload, the more that it can simply be on by default in a lot of use cases without feeling like you're giving up a lot of performance, creative freedom or accuracy. You can't have a high number of false positives or people would rather take control back.

We need local content moderation to be cool, efficient, accurate and configurable so that people can get the higher quality results which unsafe models tend to produce, while filtering out the unsafe results fast enough that it can even be done in an agentic way and fed back into the model to steer its results.

Can you describe more in terms of what moderation you want? I can say with confidence that we don't train for the typical adult-only content you can see on some "de-safe'd" models, but it's not clear to me exactly what you're asking about. We can also discuss over email if you prefer.

I'm largely referring to the fundamental components that can be combined to arrive at a decision. Is this a real life image or text, or is it some sort of art or fiction? Is this in a context of historical events for education, etc? Is this sarcasm? Is this a real threat? Is this nudity? Does this contain violence and gore? Is this suggestive? How old is this character? Is this content sufficiently related to the content it's replying to, or is it spam? Does this response deviate too much from the input request? Is this a business logo?

There was this book called The Inquisitive Mood and it very much reminds me of the chain interrogation process you can do with models now.

This isn't exactly a single use case, it's more like there are use cases for image editing software that use generative models, there are use cases for local LLMs to get a sense for abstract user intent, or to check if a model's action is in line with policy and so on.

As I said, the Liquid AI models are already very good at a lot of this, but considering how good the models are at it, I'm curious how much y'all train on tasks that are moderation-adjacent. If y'all don't have any model that's really specifically trained for these kinds of tasks, it would be interesting to see how close y'all could get if you tried.

This is an actually important problem in the AI space especially for local development.

There are other moderation tasks that are harder, like "is this copyrighted?" or "is this a McDonald's logo?" because of how much copyrighted material there is and how many businesses there are, but it might be possible to get a good feel for "is this likely a trademark?", "is this a business logo?", "does this text seem to describe a popular character?"

To some extent you can combine a model like this with a good database to optimize the filters and identification. Then you can provide a few reference image or text examples that are closer to what you're trying to filter out and the model can say, "yeah, this content is pretty close to your examples."

Google and OpenAI have already released some of their open source safety models and they were really quite bad, from my testing. Yet, Liquid AI hasn't really leaned into promoting their models for moderation purposes despite being excellent. I know there is a lot of political intersection in moderation, especially in the past many years, but I think a model that is smart enough and excels at letting people apply their own rules with fairly good accuracy is important right now.

on moderation, we run an open guardrail benchmark at raxIT Labs and the hosted d1 models do well. d1.1 is #1 of 30 at checking model output for harmful content (94.4 balanced accuracy), d1-ultra is #1 on harmful user requests (88.2). we haven't run the new 2B open weights yet so no number for those. results and method: https://dm-as-guardrails.raxitlabs.com

This tracks with my experience with their models. Nice work. It would be interesting to see what they could do if they optimized even more for moderation like this without going so far as to hurt the model performance overall.

Was there any intent behind the audio encoder representation or was it frozen backbone -> co adaptation?

speed :)

broadly, Liquid focuses on picking architectures that have really high throughput/low latency that are close to full attention (slow) in terms of expressivity, and then we do the data work to close any remaining gap. so the big goal is to make it very very fast to encode data without any loss in quality.

Agentic coding numbers on par with Qwen3.5-4B, the rest of the numbers on par with Qwen3.5-9B. Most impressive for a 1.7B model!

Liquid has also been very good about keeping speed in mind, and the writeup here makes me expect similar. Really good stuff.

Is anyone going to seriously use a model this small for coding? Doesn't seem like the intended use case or strength of these models.

Agreed not, but I think we are also moving towards a place where harnesses have code mode built in and where data swizelling and tool use is something the LLMs have to be competent at. I think that's present enough here!

And to beat Qwen3.5-4B with such a capable model is amazing, on top of so many win/draws against 9b in the more general arena.

I can definitely see there being some roles for local agents that offload shell scripting & doing some scut work. I don't know what speed to expect but like, if someone has a decent NPU or GPU and can just get a bunch of basic systems work done with a small local model that could really speed up a lot of workflows that I see day after day after day.

definitely would not recommend using it for SWE-style tasks. But chaining a couple bash calls together to find a file? no problem

No, especially not for generative coding like all folks do with Claude Code - this model is not meant for that. It will produce code that mixes too many things at once, creating spaghetti code and likely having difficulty producing a compilable artifact. I suppose tool calling is also far from the frontier of class models.

However, after testing it locally, in my imagination I find this model to be a good companion for providing TAB suggestions or essentially an "ask-a-friend" that can generate example snippets of the commonly used libraries (e.g., how to create a pivot table in pandas) or bash programs (ffmpeg command to stream internet radio).

IMHO, expect max capabilities and quality of answers that frontiers models were doing back in late 2024.

Just tried both the Omni model and the text only model, both as mlx versions, in lm studio omlx and rapid mlx, none of them could actually fire the model up, kept saying safetensors cant be read (yes I realize Claude should hear this not this thread).

I think mlx-vlm should support both the Audio and the Omni model: https://github.com/Blaizzy/mlx-vlm/ and there should be native MLX exports for both here:

- https://huggingface.co/LiquidAI/LFM3-Audio-2B-MLX-8bit

- https://huggingface.co/LiquidAI/LFM3-Omni-2B-MLX-8bit

I’ve been using 2.6B as my main home assistant model and it’s been good. I switched to Ling 3.0 Tiny since it’s faster with A1B. Really hoping 1.7B is good enough (2.5 version was too dumb for HA) so I can run everything on one machine.

what does your home assistant do? a general siri/alexa-style model? or something else?

He probably means the Home Assistant software at home-assistant.io

As mentioned by another commenter, I mean Home Assistant. Essentially the main orchestrator for all my smart home things but it doesn’t limit to just smart home. You can add calendar and location sharing to HA.

Post not up yet, but the models are: https://huggingface.co/collections/LiquidAI/lfm3

Title is: Introducing LFM3: Every Modality for Specialized Intelligence

updated, ty

Looks really neat. I'm keen to try this on my clicks communicator when that arrives.

It's mediatek NPU claims to handle ONNX. That plus the 12GB ram makes the wonder if I can get this phone to be a mini-agent.

Surprisingly (or maybe not?), LFM3-VL-2B gave me better results than LFM3-1.7B on text-only prompts (general knowledge, chatting) through liquid playground. As an example, when asked “what does a cow say?”, LFM3-1.7B said that a cow can say “Oink”, and when I asked the same question to LFM3-VL-3B, it (correctly) said “moo”. I assume this is lower quant problem, though: Q8 versions I’ve played with through llama both give the correct “moo” answer.

quantization is always a problem; what quant were you using?

Q8_0 on my laptop, hope LiquidAI team can give some feedback on what model they offer in the playground!

the playground is bf16 or fp8 depending on the cloud GPU for the playground. I didn't have any issues with cows saying moo but I suspect it's just LLM non-determinism :)

What do they say if you ask what cows drink? ;) Actually I’d be curious whether they would justify ‘milk’ with ‘when they’re very young’.

I'm testing this out as a dumb scout model alongside Qwen 3.8 Flash Next (IQ3_S, strata). Currently have it labeled as a "dumb but fast scout, findings should be verified" since some initial testing shows useful-but-sometimes-inaccurate.

on a 7900XT (20GB VRAM, ~800GB/s IO) w/ full context:

    - 8500 tok/s prefill/ingestion
    - 212 toks/s decode/generation

First class hardware support makes this really exciting and different. On device processing is cheaper and faster, critical for real time applications of AI. Really excited for the rise of local, smart intelligence!

llama.cpp looks to be cloning openai's realtime websocket API with this audio model as its first use case? I guess this model is either the first speech-to-speech of its kind, OR it's the first good-enough one. I'm thinking I need to give this more testing!

  - https://github.com/ggml-org/llama.cpp/pull/31287
  - https://github.com/ggml-org/llama.cpp/pull/31269

Loved playing with LFM2.5 and excited about this!

has anyone tried finetuning one of these yet?

Reply on news.ycombinator.com