← Back to search
#127 Robin: This 1.7B Local AI Runs Agents, Vision, Voice - Even on an $80 Arduino
AI Fire Daily · 2026-10-29 · 17 min
Show full episode description
Got something to say to this podcast? Send a voicemail, and you might hear yourself on the next episode! Liquid AI’s LFM3 family shows how far tiny local models have come. A 1.7B-parameter model can now code, edit files, run lightweight AI agent tasks, understand images and voice, and even run fully offline on low-cost hardware. We’ll talk about: LFM3 1.7B: How a tiny local model handled coding, file editing, research, and agent workflows on a regular laptop. Vision + OCR: Where LFM3 did well with images, manuals, and extracting structured information. Offline Voice + Video: What worked with audio and multimodal understanding — and where hallucinations still showed up. $80 Arduino Test: How a 4-bit LFM3 model ran on just 4GB of RAM at around 3 tokens per second. Who It’s For: Why small local models could make private AI agents and offline assistants much more accessible. Keywords: Liquid AI, LFM3, Local LLM, LFM3 1.7B, Offline AI, AI Agents, Hermes Agent, Arduino Uno Q, Edge AI, Local AI, Multimodal AI, Vision Model, Voice AI, Small Language Models. Links: Newsletter: Sign up for our FREE daily newsletter. Our Community: Get 3-level AI tutorials across industries. Join AI Fire Academy: 500+ advanced AI workflows ($14,500+ Value) Our Socials: Facebook Group: Join 300K+ AI builders X (Twitter): Follow us for daily AI drops YouTube: Watch AI walkthroughs & tutorials
✨ Episode Outline — click any point to jump to it in the episode
Problem solved
Can Liquid AI's tiny 1.7B LFM3 models run useful agent, vision, voice and omni tasks fully offline on a CPU laptop and even an $80 Arduino board.
Benefits
- Runs entirely offline on CPU with no GPU, at Q8 quantization
- 256k context window lets it ingest whole codebases or manuals locally
- Official GGUF quantizations make local deployment easy
- Hermes agent loop lets it recover from errors instead of crashing
- Strong OCR: turns blurry manual photos into clean tables
Use cases
- Built a working browser OS, then via Hermes added a taskbar, settings menu and calculator (no number pad)
- Hermes agent found a deliberately moved hidden file by searching the directory tree
- Hit permission-denied writing outside workspace, read the error and saved to a scratch location instead
- Web-researched and wrote a Wikipedia-style personal profile with references
- Vision model converted a blurry car fuse manual photo into a readable table
KPIs / results
- 1.7B parameters, trained on 22 trillion tokens
- 256k context window
- Test laptop: Microsoft Surface, 64GB unified memory, CPU-only at Q8
- Target board costs $80
Tools / build
- Liquid AI LFM3 (text, vision, audio, omni, decision models)
- Hermes agent framework
- GGUF quantizations
- $80 Arduino board
- Offline voice assistant with voice activity detection
[SPEAKER_03] A tiny AI was asked to give a browser OS more pizazz, but it completely misunderstood the assignment. It decided to build a functional pizza browser OS instead. Seriously. [SPEAKER_00] It is a genuinely hilarious mistake. But here is the truly crazy part. That tiny 1.7B parameter model did actual real work. It built working interfaces. It edited system files autonomously. And it did all of this on a standard, everyday laptop. [SPEAKER_03] Welcome to today's Deep Dive. We are exploring Liquid AI's new LFM3 family. Specifically, we are looking at their 1.7B parameter models. [SPEAKER_00] Which are incredibly small by today's standards. [SPEAKER_03] Exactly. We want to know what happens when you ditch expensive GPUs. We are running AI entirely locally, completely offline. We even test it on an $80 Arduino board. [SPEAKER_00] It is absolutely wild to think about. We are putting highly capable AI on budget hardware. That is computing power cheaper than a good pair of sneakers. [SPEAKER_03] Yeah, it really is. [SPEAKER_00] The accessibility implications here are truly massive for everyone. [SPEAKER_03] But before we give this AI eyes and ears, we need to test its raw brain power on a normal laptop. [SPEAKER_00] Right. So the LFM3 family includes five distinct models. They handle text, vision, audio, omni, and decision tasks. Mm-hmm. And each model sits right around that 1.7B parameter mark. That is a tiny footprint. [SPEAKER_03] The text model handles the heavy logical lifting, but a few specific specs really caught my eye here. It was trained on 22 trillion tokens. [SPEAKER_00] Which is just a staggering number. [SPEAKER_03] It really is. Let's define a token quickly for you. A token is basically a tiny piece of a word. That is a massive amount of data for something this small. [SPEAKER_00] It completely flips traditional machine learning training on its head. Usually developers train small models on much smaller data sets. [SPEAKER_03] Right. They scale the data to the model size. [SPEAKER_00] Exactly. But Liquid AI basically overtrained this tiny neural network on purpose. [SPEAKER_03] Why would they force so much data into a tiny model? [SPEAKER_00] Well, it forces the model to compress the underlying logic severely. It is like boiling down billions of books into a thick, dense syrup. [SPEAKER_03] Ooh, I like that analogy. [SPEAKER_00] Yeah. You keep the pure concepts, but you lose the exact verbatim text. [SPEAKER_03] That makes a lot more sense. It also has a 256k context window, which is the AI's short term memory limit during a single chat session. [SPEAKER_00] And handling that much memory usually requires giant corporate servers. [SPEAKER_03] Right. And seeing that on a local model is staggering. You can feed it entire software code bases at once. [SPEAKER_00] Yeah. Or it can read massive instruction manuals in a single prompt. They also provide official GGUF quantizations for developers. [SPEAKER_03] And GGUF is just a file format. It makes running local models much easier. But we should mention one major catch here. [SPEAKER_00] The licensing, right. [SPEAKER_03] Yeah. The LFM3 uses a proprietary license structure. It is not open source like MIT or Apache. [SPEAKER_00] Commercial developers definitely need to check those terms carefully. But let's look at the actual testing environment. The test used a standard Microsoft Surface laptop. [SPEAKER_03] Just a normal everyday work laptop. [SPEAKER_00] Exactly. It had 64 gigabytes of unified memory running on the CPU. [SPEAKER_03] Unified memory just means the CPU and graphics share the same RAM pool. But this was running strictly on the CPU, right? [SPEAKER_00] Yep. [SPEAKER_03] There was no heavy graphics card involved at all. [SPEAKER_00] Right. And CPU processing is typically much slower for artificial intelligence. But memory bandwidth matters much less for these tiny models. We ran the text model at Q8 quantization. [SPEAKER_03] Let me jump in there. Quantization means shrinking the model's file size to run on slower, cheaper hardware. [SPEAKER_00] Exactly. And Q8 is a high quality compression level. It shrinks the model without making the AI too stupid. [SPEAKER_03] Right. So basic chat responses were fast and surprisingly fluid. But we wanted to push it much harder. [SPEAKER_00] Yeah. We really wanted to break it. The first coding test was building a basic browser OS. [SPEAKER_03] The final interface was pretty simple, but the buttons worked correctly. That is impressive for a CPU only local model. [SPEAKER_00] It really is. But then things got significantly more complex. We tested it using the Hermes agent framework. [SPEAKER_03] This tool lets AI interact directly with your computer's files. [SPEAKER_00] Right. And this is where it gets really interesting. It doesn't just output text onto a screen anymore. It actually executes commands in your terminal directly. It becomes an active participant in your operating system. [SPEAKER_03] So first, the LFM3 successfully hunted down specific hidden files. It found a file after it was intentionally moved to the desktop. [SPEAKER_00] Which is huge. That requires the AI to actively search the directory structure. [SPEAKER_03] Yeah. It is not just guessing. [SPEAKER_00] Right. Next, it started editing the code it had just written. It read its own browser OS code base independently. [SPEAKER_03] And it successfully added a functional taskbar and a settings menu. [SPEAKER_00] It even built a working calculator interface from scratch. [SPEAKER_03] Though it built the calculator completely without a number pad. Right. Which is hilarious. But it highlights a specific AI weakness. Spatial reasoning in code is historically very hard for text-based models. [SPEAKER_00] Yeah. They struggled to visualize the layout. [SPEAKER_03] And of course, it confused pizazz with pizza. I still wrestle with prompt drift myself. So seeing this tiny model handle autonomous agent tasks is wild. [SPEAKER_00] The autonomous looping is the actual magic part here. The agent architecture forces the AI to observe, act, and evaluate. [SPEAKER_02] It tries something, checks the result, and adjusts its approach. [SPEAKER_00] Exactly. Let's look at the system permission test as an example. It tried to write a file outside its designated workspace. [SPEAKER_03] The operating system naturally threw a permission denied error. A normal script would just crash and give up right there. [SPEAKER_00] But this agent read the error. It understood the problem. Yeah. It smartly saved the updated file to a permitted scratch location instead. That shows a level of dynamic problem-solving you rarely see. [SPEAKER_03] It also conducted independent online research perfectly. It used web searches to gather external information. [SPEAKER_00] Right. It created a Wikipedia-style personal profile with references. [SPEAKER_03] Some of the personality analysis was definitely a bit questionable. But the agent workflow functioned exactly as intended. [SPEAKER_00] So, does a 256k contacts window even matter if the model isn't smart enough to use it? Well, having massive memory doesn't guarantee deep comprehension. It will recall every detail you provide flawlessly. But if the instruction is too nuanced, it will execute the wrong logic entirely. [SPEAKER_03] It remembers everything perfectly but might completely misunderstand your actual complex instructions. Got it. [SPEAKER_00] Exactly. It has the data, just not always the wisdom. [SPEAKER_03] Since we know the text model struggles with nuanced instructions, let's change gears. What happens when we force it to look at the real world? [SPEAKER_00] Multimodal capabilities make local LLMs vastly more useful. We tested the vision model with four real-world images. [SPEAKER_03] We deliberately chose tricky visual scenarios to test its limits. [SPEAKER_00] Right. The first test was a photo of a Westie dog. [SPEAKER_03] The AI correctly identified the specific dog breed, but it completely missed the emotion of the animal. The dog was clearly upset in the photo. [SPEAKER_00] Yeah, but the AI thought the dog looked completely relaxed. It failed to read the nuanced emotional context. [SPEAKER_03] Why does that happen? I mean, it can identify the breed perfectly. [SPEAKER_00] Well, vision models map geometric patterns and textures. They learn the pixel arrangement of a Westie's fur. [SPEAKER_03] Right, the physical shape. [SPEAKER_00] Exactly. But they do not possess actual empathetic context. Recognizing a shape is easy. Interpreting a feeling requires complex human intuition. [SPEAKER_03] That makes perfect sense. It sees the pixels, not the mood. [SPEAKER_00] Yep. The second image was a Subaru WRX meme. The model understood part of the visual joke. [SPEAKER_03] But it completely missed the head gasket humor. [SPEAKER_00] Right. That requires highly specific automotive cultural knowledge. [SPEAKER_03] Yeah. Multimodal models struggled to cross-reference image recognition with niche cultural text data. [SPEAKER_00] It recognized the car, just not the internet joke associated with it. [SPEAKER_03] Next, we showed it a Toshiba Cosmio laptop. It recognized the laptop by reading its label. [SPEAKER_00] But it actually misspelled the name in its final output. [SPEAKER_03] It also struggled heavily with the tiny chassis stickers. It could not decipher the small logos accurately. [SPEAKER_00] Small text in blurry images is really tough for a tiny parameter model. [SPEAKER_03] But the final image was a car fuse manual. This result was actually incredibly useful and practical. [SPEAKER_00] Yeah, this was a huge win. It converted a blurry photo into a clean, readable table. [SPEAKER_03] Imagine taking a photo of a dense manual with your phone. [SPEAKER_00] Exactly. You can have an offline AI format it instantly. Optical character recognition is a major strong point for this model. [SPEAKER_03] Reading a fuse box perfectly but missing a dog's emotion is functionally weird. [SPEAKER_00] Right, and it shows how these models actually see the world. They are excellent at finding rigid patterns like formatted text. But they struggle with fluid concepts like animal emotions or humor. You still need to double check important details constantly. [SPEAKER_03] Since it struggles with visual subtext, I'm curious about audio. Does that lack of nuance bleed into how it processes human voices? [SPEAKER_00] Well, the audio model functions as a voice assistant without internet. It features built-in voice activity detection technology. [SPEAKER_03] This means it listens and responds automatically to your voice. You do not have to press a button each time you speak. [SPEAKER_00] Right. The conversation flow felt incredibly quick and snappy. But accuracy was an entirely different story. [SPEAKER_03] Yeah, it suffered from some very confident hallucinations. It completely invented specifications for a MacBook Pro. [SPEAKER_00] It claimed the laptop had Apple Pencil support. [SPEAKER_03] It even confidently made up a fake phone number. It presents false information with absolute unwavering certainty. [SPEAKER_00] That happens a lot with these small models. [SPEAKER_03] But why does an offline model just invent a phone number? [SPEAKER_00] When an offline model lacks internet grounding, it defaults to statistical probability. [SPEAKER_03] Okay, what does that mean in practice? [SPEAKER_00] It knows what a phone number looks like structurally. So it just generates a sequence of numbers that fits the pattern. [SPEAKER_03] Ah, I see. [SPEAKER_00] It does not know the number is fake. It just knows the format is right. [SPEAKER_03] That is slightly terrifying, honestly. It also severely misunderstood human speech during testing. [SPEAKER_00] Yeah, when asked to draft a simple message, it failed. It completely reversed the meaning of a key sentence. [SPEAKER_03] It also lacked a vital conversational feature. It could not be interrupted while it was speaking. [SPEAKER_00] Right. You literally have to wait for it to finish its entire thought completely. [SPEAKER_03] That makes the interaction feel incredibly rigid and unnatural. Humans interrupt each other constantly for clarification. [SPEAKER_00] We really do. Still, having a working offline voice assistant is an impressive baseline. We also tested the OmniModel capabilities. [SPEAKER_03] The OmniModel processes text, images, video, and audio simultaneously. [SPEAKER_00] Right. And it can respond using either text or synthesized speech output. We tested it on a short promotional video clip. [SPEAKER_03] It was an advertisement for the open source Perry chatbot. The model responded surprisingly fast to the video feed. [SPEAKER_00] It understood most of the visual content perfectly, but it completely misread the text saying to join the waitlist. [SPEAKER_03] Also, the audio understanding completely failed in this specific test. [SPEAKER_00] Yeah, but the audio failure was likely a web UI issue though. The custom interface simply did not pass the audio correctly to the model. [SPEAKER_03] But the core video understanding was highly solid for its tiny size. So, is an offline voice assistant actually useful if it confidently invents fake phone numbers? [SPEAKER_00] Well, it depends entirely on what you ask it to do. It is fantastic for controlling local devices where facts don't matter, but you shouldn't use it for research or data retrieval. [SPEAKER_03] Great for offline smart home commands, but terrible for factual research or heavy tasks. That is a very fair assessment. Okay, so if LFM3 runs decently on a normal laptop, how far down the hardware ladder can we push it? [SPEAKER_00] Right. Let's completely ditch the PC. Let's see where this technology actually breaks. [SPEAKER_03] This is where things get truly fascinating. Can a 1.7 billion parameter AI run on an $80 microcomputer? [SPEAKER_00] Can it do this entirely without internet access or massive servers? [SPEAKER_03] The answer is surprisingly yes. [SPEAKER_00] We tested it on an Arduino Uno Q setup. [SPEAKER_03] This tiny board uses a Qualcomm Dragon Wing chip. [SPEAKER_00] Let's explain what that hardware actually is. The Qualcomm Dragon Wing is a highly efficient mobile processor. [SPEAKER_03] It is designed specifically for low power robotics and edge devices. [SPEAKER_00] And it is paired with just 4 gigabytes of LPDDR4 RAM. [SPEAKER_03] That is significantly less memory than a cheap smartphone. [SPEAKER_00] Oh, way less. Let's break down the exact hardware costs involved here. The Arduino board itself costs $80. [SPEAKER_03] You also need a $10 USB-C adapter. [SPEAKER_00] And you need a $20 power brick to run it. The total cost comes out to $110. [SPEAKER_03] You will need a standard monitor and a keyboard too, of course. [SPEAKER_00] Right. But you can reuse old peripherals you already have sitting around. [SPEAKER_03] To make the model actually fit, we use quantization again. We used an extreme 4-bit quantization for this specific test. [SPEAKER_00] This level of compression is absolutely essential for devices with limited RAM. [SPEAKER_03] Yeah. It drastically reduces the mathematical precision of the model's internal weights. It essentially chops off the decimal points to save space. [SPEAKER_00] Right. But how fast does it actually run on the board? [SPEAKER_03] The Arduino manages around 3 tokens per second. A token is roughly three-quarters of a standard English word. [SPEAKER_00] That is undeniably slow for a long, complex conversation. It feels like watching someone type very, very methodically. [SPEAKER_03] It is painfully slow for real-time human chat. But the incredibly important part is that it actually works. [SPEAKER_00] Well, imagine scaling this to a billion cheap household edge devices. [SPEAKER_03] That is a staggering thought. You are putting localized brains into everyday dumb appliances. [SPEAKER_00] Yeah. [SPEAKER_03] Six years ago, this concept would have sounded completely crazy. Running functional AI on an $80 computer was pure science fiction. [SPEAKER_00] And now it is literally sitting on a desk working. Yes, you sacrifice raw speed. And you lose some intelligence. [SPEAKER_03] But you gain a private, entirely offline AI assistant. [SPEAKER_00] Exactly. It never sends your data to a corporate server. [SPEAKER_03] You avoid expensive hardware purchases entirely. [SPEAKER_00] You also eliminate all ongoing cloud API costs forever. [SPEAKER_03] But at a painfully slow three tokens a second, isn't this just a frustrating tech novelty? [SPEAKER_00] Well, when you are building a smart home system, latency is secondary. If your goal is zero data leakage, you gladly accept the slower speed. [SPEAKER_03] Speed doesn't matter when your priority is absolute privacy and zero ongoing cloud costs. [SPEAKER_00] Exactly. [SPEAKER_03] That makes perfect sense for many users. The latency is a fair trade for complete digital sovereignty. Let's synthesize the main takeaway for you today. [SPEAKER_00] Liquid AI's LFM3 isn't trying to dethrone massive cloud models. It is not competing with GPT-4 or CLOD directly. [SPEAKER_03] It serves an entirely different purpose in the AI ecosystem. Right. [SPEAKER_00] It is proving that highly capable private AI is now accessible. [SPEAKER_03] Everyday people can run it on standard, affordable hardware. We must acknowledge the current warts and flaws, obviously. [SPEAKER_00] Yeah, we saw the confident hallucinations with the fake phone numbers. [SPEAKER_03] And the optical character recognition is still imperfect on tiny stickers. [SPEAKER_00] The coding results are sometimes clumsy, like the missing numpad. [SPEAKER_03] Right. And the proprietary licensing is a drawback for commercial developers. [SPEAKER_00] Sure. But we absolutely cannot let those flaws obscure the bigger picture here. We are witnessing a monumental shift in computing infrastructure right now. [SPEAKER_03] Running a functional AI chatbot on an $80 offline board changes everything. [SPEAKER_00] It completely decentralizes where artificial intelligence can live and operate. [SPEAKER_03] It removes the massive financial barrier to entry for experimentation. You do not need thousands of dollars in server equipment anymore. [SPEAKER_01] You just need curiosity and a basic laptop computer. Or just $100 and a tiny Arduino board. [SPEAKER_03] The democratization of this technology is happening incredibly fast. These small, highly efficient models are going to become ubiquitous very soon. [SPEAKER_00] Absolutely. [SPEAKER_03] We want to leave you with a final unmentioned thread to mull over. Six years ago, this level of local AI was unthinkable. Right. If an $80 board can act as an autonomous offline agent today, what happens to our privacy and daily routines when every appliance in your house has its own localized 1.7 billion parameter brain quietly making decisions in the background? Thank you for joining us on this deep dive. [SPEAKER_03] We appreciate your time and your curiosity today. Take care and keep asking the big questions. Take care and keep asking the questions. You go ahead of luck if the virus of a child may beakovany or STOP were able to connect to darknessagn Ghost globalDB