This is cool. I definitely think the "micro" sized LLM space is underappreciated, so it's always good to see work like this. I foresee a paradigm in some contexts where you have a hierarchy of LLMs, with more competent models actively training smaller models to solve specific tasks very efficiently, and something like this could be the smallest layer in that stack.
With that being said, the web demo is not particularly impressive. It really doesn't like anything I throw at it. I'm fine with accepting that fine-tuning is the solution to this, but I wonder if there's anything to gain from a bigger model? I know it's completely counter to the whole point of this, but a 14MB binary using 28MB of RAM seems unnecessarily small and pretty arbitrary.
Like, what does a 28MB binary get you? Or a 140MB binary? Or a 1.4MB binary? I'm guessing the choice of 14MB came from minimizing the size as much as possible while meeting certain requirements/performance expectations, but even a Pi 5 has plenty more room to spare. Curious if there's a good explanation for this (which I may have missed in my skim of the post).
So, its not a general language model, focused on tool call strictly for tiny edge-devices. There are solutions everywhere for high-capacity devices, Needle is for sub-$200 devices.
Funny result from the web demo. I'm well aware that it's an extremely small and, well, stupid, model, but even so:
Query: HN
Result:
{
"function_calls": [
{
"name": "lock_door",
"arguments": {
"door": "front door"
}
}
],
"reasoning": "User wants to lock the door. No specific door mentioned, so use 'front door' as default.",
"confidence": 0
}
I'd expect it to at least ignore (call no tools) for the queries that it doesn't understand. And it seems like it does do that, just not consistently.
The website says the model is for "tool calling, device use, and structured extraction". Your example just doesn't seem to be very relevant. FWIW, it did a pretty good job for tool calling when I tried it, and I think it could be pretty nice to have this running on locally and integrate with Home Assistant.
False positives are definitely relevant and worth measuring - natural language interfaces always have a discoverability problem, i.e., users not knowing what actions the system does and does not support. If the frontend of that system lacks the ability to reject unsupported commands, weird stuff happens.
Nonetheless, this is very cool work! If I can offer a small suggestion to the team at Cactus, it would be to evaluate your releases on some usability criteria (including false positives). Any serious integrator or adopter of these models would want to have that information available.
Right, but that's not the same thing as reporting a benchmark across a test set. It doesn't help me determine how well the model does across a decently-large sample size of commands. It doesn't tell me with what reliability the confidence will be below a given threshold when it should be, above that threshold when it should be, etc.
Makes me think of the demo from some time ago where someone got a ~29M parameter model running on an esp32. I wonder what kind of throughput this could get if a handful of esp32s were strung together...
That's really cool - I was already thinking of compressing `functiongemma-270m-it` down to 1-2 bits so it would work flawlessly in the browser.
Your `Fine-tuning` feature is even much more convenient.
While most of the industry focuses on the frontier of “intelligence” (function), a release like this represents the frontier of the other end of the spectrum (form).
Both are important if we ever want to see “Opus-level” capability running locally on commodity machines in the future.
This is really cool, I'm curious how much knowledge can their be in smaller models? It seems the current trade off is you need sizeably larger models for more performance but I'm curious if in your work how far this is true, as edge ai is really what needs to get better before physical ai can take off (my two cents).
I imagine at such a low parameter count, there would be little to no world knowledge whatsoever, and the entire focus is on getting the structure of tool calling etc. right…?
But yeah, in terms of “physical” AI, robotics definitely comes to mind for me as well, where tool calls/structured “device” use in a “realtime”/edge application are highly beneficial (if you wanted to go with LLMs), but beefy hardware can’t be easily used.
Was really cool to see yous use Engrams to cut down compute!
Given its basically an O(1) lookup with disk space being the main constraint, I was curious if you've tried ablating engram layers and sizes across your setup?
That will get you a lot further than what you're asking, but if you dig a bit through Home Assistant features, resources, etc., you may find the current "best" answers to your questions.
If you want a quick answer: Whisper is a good open-source speech-to-text model which comes in a variety of sizes (https://huggingface.co/openai/whisper-tiny). You can definitely get something like this running on a Pi 5. There are plenty of other STT models out there, some of which are built specifically for this context (again, see the Home Assistant stuff), but Whisper comes up a lot as a good default choice.
So with something like Whisper, you could just have a simple script which is constantly listening to a rolling window of audio and transcribing it. When the transcription includes a key phrase, you can pass the rest of the transcription to Needle2 (or anything else for that matter). From there, you take the results and execute the necessary tool calls.
There's a bit more to all of this to make it work smoothly, but fundamentally this is all there is to it. All this would work very fast on a Pi 5 (although I wouldn't expect the results to be particularly good without some serious hand-crafted logic, fine-tuning, etc.). If you want to mess around this stuff, handing all of this to Claude, Codex, etc., can get you something spun up and functional very quickly.
The learned confidence gate is the crucial piece for a 14MB action model. On ambiguous requests such as the HN example, what calibration target decides between abstaining locally and escalating to the cloud?
around +60% confidence threshold is cool from experiments, the problem is that you gotta test on your own workload, no existing benchmark could honestly paint the full picture, so we exposed the confidence threshold for everyone.
Could someone please share how such open source micro-LLMs might have been created?
Do the creators take something like DeepSeek, and then delete most of the neurons to whittle down the size?
This is cool. I definitely think the "micro" sized LLM space is underappreciated, so it's always good to see work like this. I foresee a paradigm in some contexts where you have a hierarchy of LLMs, with more competent models actively training smaller models to solve specific tasks very efficiently, and something like this could be the smallest layer in that stack.
With that being said, the web demo is not particularly impressive. It really doesn't like anything I throw at it. I'm fine with accepting that fine-tuning is the solution to this, but I wonder if there's anything to gain from a bigger model? I know it's completely counter to the whole point of this, but a 14MB binary using 28MB of RAM seems unnecessarily small and pretty arbitrary.
Like, what does a 28MB binary get you? Or a 140MB binary? Or a 1.4MB binary? I'm guessing the choice of 14MB came from minimizing the size as much as possible while meeting certain requirements/performance expectations, but even a Pi 5 has plenty more room to spare. Curious if there's a good explanation for this (which I may have missed in my skim of the post).
So, its not a general language model, focused on tool call strictly for tiny edge-devices. There are solutions everywhere for high-capacity devices, Needle is for sub-$200 devices.
I tested with
python main.py No calculator or math tool available.conclusion: completly useless
Funny result from the web demo. I'm well aware that it's an extremely small and, well, stupid, model, but even so:
Query: HN
Result:
{ "function_calls": [ { "name": "lock_door", "arguments": { "door": "front door" } } ], "reasoning": "User wants to lock the door. No specific door mentioned, so use 'front door' as default.", "confidence": 0 }
I'd expect it to at least ignore (call no tools) for the queries that it doesn't understand. And it seems like it does do that, just not consistently.
The website says the model is for "tool calling, device use, and structured extraction". Your example just doesn't seem to be very relevant. FWIW, it did a pretty good job for tool calling when I tried it, and I think it could be pretty nice to have this running on locally and integrate with Home Assistant.
False positives are definitely relevant and worth measuring - natural language interfaces always have a discoverability problem, i.e., users not knowing what actions the system does and does not support. If the frontend of that system lacks the ability to reject unsupported commands, weird stuff happens.
Nonetheless, this is very cool work! If I can offer a small suggestion to the team at Cactus, it would be to evaluate your releases on some usability criteria (including false positives). Any serious integrator or adopter of these models would want to have that information available.
> "confidence": 0
OP and the linked page talk about the confidence score and using it as an action threshold, so it looks like an appropriate total response to me.
Right, but that's not the same thing as reporting a benchmark across a test set. It doesn't help me determine how well the model does across a decently-large sample size of commands. It doesn't tell me with what reliability the confidence will be below a given threshold when it should be, above that threshold when it should be, etc.
Thanks, noted!
no, this is the appropriate response to hearing the words "HN" :P
"confidence": 0, so I guess you could threshold it
yeah I got the same, almost like its biased heavily towards that as the 0 ranking -- my prompt was just the word 'potato'
Was that the first message you sent it?
Makes me think of the demo from some time ago where someone got a ~29M parameter model running on an esp32. I wonder what kind of throughput this could get if a handful of esp32s were strung together...
That's really cool - I was already thinking of compressing `functiongemma-270m-it` down to 1-2 bits so it would work flawlessly in the browser. Your `Fine-tuning` feature is even much more convenient.
Thanks, give it a splin!
This is cool!
While most of the industry focuses on the frontier of “intelligence” (function), a release like this represents the frontier of the other end of the spectrum (form).
Both are important if we ever want to see “Opus-level” capability running locally on commodity machines in the future.
thanks!
This is really cool, I'm curious how much knowledge can their be in smaller models? It seems the current trade off is you need sizeably larger models for more performance but I'm curious if in your work how far this is true, as edge ai is really what needs to get better before physical ai can take off (my two cents).
I imagine at such a low parameter count, there would be little to no world knowledge whatsoever, and the entire focus is on getting the structure of tool calling etc. right…?
But yeah, in terms of “physical” AI, robotics definitely comes to mind for me as well, where tool calls/structured “device” use in a “realtime”/edge application are highly beneficial (if you wanted to go with LLMs), but beefy hardware can’t be easily used.
Was really cool to see yous use Engrams to cut down compute!
Given its basically an O(1) lookup with disk space being the main constraint, I was curious if you've tried ablating engram layers and sizes across your setup?
Also, why mHC over attention residuals?
Yes, we ablated Engrams rigorously and found that it returned world knowledge like FFN without without compute expenditure.
What about mHC? I'm surprised it helped with such a small compute budget.
I’d be interested in attempting to run this in a network- enabled 32MiB RAM microVM.
Thanks, how can we help?
Naïve and clumsy question: how would you pair this with speech-text-speech stuff, wake words etc.? Are there good examples of this for a Pi 5?
The demo is super — I'm just having trouble seeing the whole picture for e.g. a screenless device.
ETA: pun not intended
The best entrypoint is Home Assistant: https://www.home-assistant.io/
That will get you a lot further than what you're asking, but if you dig a bit through Home Assistant features, resources, etc., you may find the current "best" answers to your questions.
If you want a quick answer: Whisper is a good open-source speech-to-text model which comes in a variety of sizes (https://huggingface.co/openai/whisper-tiny). You can definitely get something like this running on a Pi 5. There are plenty of other STT models out there, some of which are built specifically for this context (again, see the Home Assistant stuff), but Whisper comes up a lot as a good default choice.
So with something like Whisper, you could just have a simple script which is constantly listening to a rolling window of audio and transcribing it. When the transcription includes a key phrase, you can pass the rest of the transcription to Needle2 (or anything else for that matter). From there, you take the results and execute the necessary tool calls.
There's a bit more to all of this to make it work smoothly, but fundamentally this is all there is to it. All this would work very fast on a Pi 5 (although I wouldn't expect the results to be particularly good without some serious hand-crafted logic, fine-tuning, etc.). If you want to mess around this stuff, handing all of this to Claude, Codex, etc., can get you something spun up and functional very quickly.
This is a very responsive answer, thank you so much. (I'd assumed maybe Whisper but the wake word "loop" detail there is illuminating.)
Users often stack a transcription model on top to get the voice prompt, then decode to actions. Think of Alexa and Siri.
Thank you.
The learned confidence gate is the crucial piece for a 14MB action model. On ambiguous requests such as the HN example, what calibration target decides between abstaining locally and escalating to the cloud?
around +60% confidence threshold is cool from experiments, the problem is that you gotta test on your own workload, no existing benchmark could honestly paint the full picture, so we exposed the confidence threshold for everyone.
What is the difference between this and random sentence generator?
Random sentence is not a function call.
Ask it to lock a door for instance. It seems to convert simple instructions to reasonable tool calls. Check its confidence score.