I had a thing[1] over 10 years ago that could handle this kind of problem using SPARQL and knowledge graphs.
My question is how effective is it at handling ambiguity.
Can I send it something like a text message "lets catch up at coffee tomorrow 10:00" and a command like "save this" and have it choose a "add appointment" action from hundreds (or even tens) of possible tools?
Thanks to a Huggingface linked below, I tested it and im not impressed. prmopt: i need to contact my boss i will be late. Result: 20mins [{"name":"set_timer","arguments":{"time_human":"20 minutes"}}]. It didnt use the email tool and i tried 2-3 different ways of asking it.
HnUser12 2 hours ago [-]
Did you give it an email tool? It uses the tool it’s given. HF example only has timer tool.
mahmoudimus 2 hours ago [-]
works for me:
input: i need to contact my boss i will be late.
output: [{"name":"send_email","arguments":{"to":"boss@company.com","subject":"Running late","body":"I will be late for the meeting."}}]
it did have the send_email tool on the left hand side though
Google reportedly reacts to distillation attempts "with real-time proactive defenses that can degrade student model performance". So if they detected you, they could have intentionally fed you a dumber but plausible variant of Gemini: https://cloud.google.com/blog/topics/threat-intelligence/dis...
But also, this model is small and just focusing on the tool use. In terms of token usage, you're probably not anywhere near the people that are trying to distill the entire model.
madduci 20 minutes ago [-]
Well, it's like robbing the robbers, when it comes to training data
tommica 9 minutes ago [-]
Except one of the robberers is a massive corporation with even bigger legal team...
jumploops 29 minutes ago [-]
This is neat, and matches an observation I saw with early Claude Code usage:
Sonnet would often call tools quickly to gather more context, whereas Opus would spend more time reasoning and trying to solve a problem with the context it had.
This led to lots of duplicated functions and slower development, though the new models (GPT-5.5 and Opus 4.6) seem to suffer from this less.
My takeaway was that “dumber” (i.e. smaller) models might be better as an agentic harness, or at least feasibly cheaper/faster to run for a large swath of problems.
I haven’t found Gemini to be particularly good at long horizon tool calling though. It might be interesting to distill traces from real Codex or Claude code sessions, where there’s long chains of tool calls between each user query.
Personally, I’d love a slightly larger model that runs easily on an e.g. 32GB M2 MBP, but with tool calling RL as the primary focus.
Some of the open weight models are getting close (Kimi, Qwen), but the quantization required to fit them on smaller machines seems to drop performance substantially.
ilaksh 10 hours ago [-]
Hmm.. this might make it feasible to build something like a command line program where you can optionally just specify the arguments in natural language. Although I know people will object to including an extra 14 MB and the computation for "parsing" and it could be pretty bad if everyone started doing that.
But it's really interesting to me that that may be possible now. You can include a fine-tuned model that understands how to use your program.
E.g. `> toolcli what can you do` runs `toolcli --help summary`, `toolcli add tom to teamfutz group` = `toolcli --gadd teamfutz tom`
HenryNdubuaku 10 hours ago [-]
So Needle is trained for INT4, what you see in the playground is INT4, only 14MB, same challenge though.
ilaksh 10 hours ago [-]
Oh gotcha. Fixed my comment.
simonw 10 hours ago [-]
Suggestion: publish a live demo of the "needle playground". It's small enough that it should be pretty cheap to run this on a little VPS somewhere!
quantumleaper 9 hours ago [-]
Should be quick and easy with WebGPU, too.
simonw 9 hours ago [-]
That's an even better idea, I bet this could run in Transformers.js.
ilaksh 9 hours ago [-]
Good idea. Could you make that.
bijowo1676 5 hours ago [-]
Good idea. Could you ask a Claude Code to make that.
Today is 2026 after all
utopiah 7 minutes ago [-]
It's 2026 so it's already been done 10x by 5x people who says AI is amazing but none of them is sharing the outcome because they either don't care or it doesn't even work.
HenryNdubuaku 10 hours ago [-]
thanks, yeah, the problem is just handling scale, we don't have the infra ready to go, but anyone can do that. Its easy for people to run on their laptops straight up. Will try the VPS route.
I know we all think of bad things when we hear "short form video" but short demos can do a LOT for any project, shows the user how its used, what it looks like, what it solves, etc all in anywhere from 15 seconds to a couple of minutes, doesn't need to be ultra fancy, screen recording is fine. :)
bityard 8 hours ago [-]
Since there is no GUI here, I feel like a simple plaintext chat transcript would be both 100x smaller and 100x easier to read. (Not to mention accessible.)
giancarlostoro 8 hours ago [-]
Sure, and we've seen those terminal screen recorders that give you back a replayable demo, that could work too.
Barbing 1 hours ago [-]
One of the most important things missing from too many projects. Even fifteen seconds can often help significantly.
HenryNdubuaku 8 hours ago [-]
Yes, a demo might be a good idea.
bilalba 3 hours ago [-]
I'll put this on chonklm.com!
kristopolous 9 hours ago [-]
That M versus B is way too subtle. 0.026B is my suggestion
bigyabai 5 hours ago [-]
The "M" nomenclature has been around since at least BERT and T5/FLAN. It's valid to use it even if today's LLM devs are more familiar with billion-scale models.
DrammBA 3 hours ago [-]
I was so confused by many comments in this post but thanks to you I realized that some people are apparently reading it as 26B and that's why their comments make no sense.
HenryNdubuaku 9 hours ago [-]
Haha, we were trying to not be hand-wavy too much :)
kristopolous 3 hours ago [-]
Oh hey it's Henry. I met you a couple weeks ago at an event in SF. Nice to see you on here.
9 hours ago [-]
dymk 7 hours ago [-]
[flagged]
dang 3 hours ago [-]
Can you please make your substantive points without sharp elbows? We're trying for something different here, and would appreciate it if you'd post in the intended spirit.
I’d edit it if I could, but it seems to be past the timeout.
As the other poster noted, the post wasn’t meant to be read as a personal attack
dang 55 minutes ago [-]
I've reopened it for editing if you want to (it's totally fine either way - we just care about fixing things going forward)
kristopolous 6 hours ago [-]
Pardon me, do I know you?
Why are you attacking me?
osrec 5 hours ago [-]
I don't think they're attacking you, but suggesting you read more carefully. The information provided is correct and clear, but you need to let go of your own biases when consuming it.
I personally prefer the M to the B. I guess as an engineer, noticing the units comes pretty naturally.
kristopolous 4 hours ago [-]
25-35 Billion is expected these days, there's many models of this size, it's very common. (Gemma 4 31B, Qwen 3.6 25B & 35B, JT 35B, EXAONE 35B, Nemotron 30B, GLM 4.7-flash 30B, Servam 30B, LFM2 24B, Granite 4.1 30B...)
Announcing something that's 1/1000th is significant and remarkable! Hiding it in a single letter is burying the lede.
f33d5173 6 hours ago [-]
I read it as 26B as well.
tomaskafka 7 hours ago [-]
Awesome! I just tried to set an alarm and add some groceries to the shopping list, and it outperformed Siri.
HenryNdubuaku 7 hours ago [-]
Music to our ears!
brainless 6 hours ago [-]
Lovely to see the push for tiny models.
I have been building for small (20B or less) models for quite a while. Highly focused/constrained agents, many of them running together in some kind of task orchestration mode to achieve what feels like one "agent".
I build (privacy first) desktop apps this way and I want to get into mobile apps with similar ideas but tiny models.
HenryNdubuaku 5 hours ago [-]
Give it a go and let us know!
kgeist 5 hours ago [-]
>Experiments at Cactus showed that MLPs can be completely dropped from transformer networks, as long as the model relies on external knowledge source.
Heh, what a coincidence, just today one of my students presented research results which also confirmed this. He removed MLP from Qwen and the model still could do transformation tasks on input but lost knowledge.
exabrial 7 hours ago [-]
Dumb questions, from someone not in the field...
What is a distilled model?
Why doesn't Google do this (to make their models smaller)?
Seems like you could make a competitor to Gemini?
HenryNdubuaku 7 hours ago [-]
No question is stupid!
1. Distilled means taking the intelligence of a big model and compacting into a tiny model.
2. Google already does so with FunctionGemma, but Needle argues that better performance could be achieved with 10x smaller model using our technologies.
tintor 6 hours ago [-]
Model distillation is lossy compression of big model to produce a smaller model.
Smaller model requires less space on disk, less video memory, and less compute (cheaper hardware).
Downside is that distilled model performs worse on the same benchmarks compared to original model.
Got a bunch of errors trying to run it on CPU though. Very likely connected to me running this in a container (unpriv LXC), but figured for 26M CPU would suffice.
It better, considering its purpose is to run on devices with no GPU.
bityard 8 hours ago [-]
This is pretty much exactly what I want for Home Assistant. I yell out, "Computer! Lights!" and it toggles the lamp in the room on or off. (I mean I can do that now, I think, but probably with a much larger model.)
I haven't played with it yet, but does it ever return anything other than a tool call? What are the failure modes? What if it doesn't understand the request? Does it ever say it can't find a tool? Does it get confused if there are two similar (but different) tools? Can it chain tools together (e.g. one tool to look up and address and another to get directions to the address)?
I mean, I plan on downloading the model later tonight and finding out for myself, but since I'm stuck at work right now, I figured I'd ask anyway...
xrd 2 hours ago [-]
Hmm, I wonder if I can run this on my MyCroft II (now NeonOS) open source AI device...
0cf8612b2e1e 6 hours ago [-]
How many lights are there?
kennywinker 5 hours ago [-]
… four. There are four lights.
HenryNdubuaku 7 hours ago [-]
Let me know what you think!
rsolva 8 hours ago [-]
Can it summarize text it fetches?
Come to think of it, this could be a nice model to have as the first pass in a more complex agent system where Needle hands of the results of a tool call to a larger model.
I will defiantly play around with this!
NordStreamYacht 5 hours ago [-]
> I will defiantly play around with this!
Are you Calvin or Hobbes?
HenryNdubuaku 8 hours ago [-]
The codebase is fully open, feel free to play around!
alex7o 7 hours ago [-]
From all the models that do toolcalls the only thing I am confused is why did you pick the worst? Or maybe they are only bad in agentic work it fine for one shot toolcalls?
HenryNdubuaku 7 hours ago [-]
Gemini is pretty solid for 1-shot tool call and affordable as well.
pylotlight 2 hours ago [-]
My general understanding of the concenus on most models these days is that people consider google models to be some of the worst at tool calling, so certainly an interesting choice. Did you do any evals on this?
BuyG1n 5 hours ago [-]
Hi, would love to know where you get that impression on 1 shot tool calling, was there concrete evaluation carried out? pretty new to this and was a bit lost when trying to compare models on different capabilities.
efskap 5 hours ago [-]
No FFN is blowing my mind. This is pretty much "Attention Is ACTUALLY All You Need". Reminds me of BERT Q&A which would return indices into the input context, but even that had a FFN. Really exciting work.
krackers 3 hours ago [-]
I guess this had always been bugging me. I get while you need activation/non-linearities, but do you really need the FFN in Transformers? People say that without it you can't do "knowledge/fact" lookups, but you still have the Value part of the attention, and if your question is "what is the capital of france" the LLM could presumably extract out "paris" from the value vector during attention computation instead of needing the FFN for that. Deleting the FFN is probably way worse in terms of scaling laws or storing information, but is it an actual architectural dead-end (in the way that deleting activation layer clearly would be since it'd collapse everythig to a linear function).
Majromax 2 hours ago [-]
> if your question is "what is the capital of france" the LLM could presumably extract out "paris" from the value vector during attention computation instead of needing the FFN for that.
But how do you get 'Paris' into the value vector in that case? The value vector is just the result of a matrix multiplication, and without a nonlinearity it can't perform a data-dependent transformation. Attention still acts as a nonlinear mixer of previous values, but your new output is still limited to the convex combination of previous values.
krackers 52 minutes ago [-]
> But how do you get 'Paris' into the value vector in that case?
Ok wait I think I see what you mean. Although maybe it's not getting paris _into_ the value vector that's hard, but isolating the residual stream to _only_ that instead of things like other capitals.
So as a naive example maybe at the very first layer consuming your tokens: Q{France} would have high inner product with K{capital} and so our residual would now mostly contain V{capital}, which maybe contains embeddings of all the capitals of all countries. You need some way to filter out all the other stuff, but can't do that without a FFN + activation.
Just throwing in a relu by itself won't help since that would still work on all the elements uniformly, you need some way to put weight on "paris" while suppressing the others, i.e. mixing within the residual stream itself.
Although maybe if you really stretch it, somewhere in a deeper layer you could have 1-hot encoded values with a "gain" coefficient so that when you do the residual addition it's something like {<paris>, <tokyo>, <dc>} + 10000*{<1>, <0>, <0>} and then if you softmax that you get something with most of its mass on "Paris". But it seems like this would not be practical, or it's just shifting the issue to how that the right 1-hot vector is chosen
murkt 9 hours ago [-]
Can this be a Siri-like core? Set me a timer, tell me what’s the weather, etc. Here is transcribed text and available list of tools for the model to call, and voice the output.
HenryNdubuaku 9 hours ago [-]
That was the goal!
z3ugma 7 hours ago [-]
I don't really understand what this is for... there is a lot of ML-researcher talk on the GH page about the model architecture, but how should I use it?
Is it a replacement for Kimi 2.7, Claude Haiku, Gemini Flash 3.1 lite, a conversational LLM for the situations where it's mostly tool-calling like coding and conversational AI?
HenryNdubuaku 7 hours ago [-]
It is for building agentic capabilities into very small devices like phones, glasses, watches and more. Does that make sense?
jcgrillo 6 hours ago [-]
I'm having trouble understanding why someone would want that? Like, what are the product use-cases of such a thing? I understand why people want that for coding agents--although the jury is still very much out on whether those are terribly useful--but I cannot fathom what someone might want an agent to do on a cell phone? Is there some user-facing activity on a phone that's similar to coding with a tight, objectively measurable feedback loop (analogous to dev/compile/test)?
EDIT: more of you cretins have downvoted than have replied.. so.. show your cards.
hosh 2 hours ago [-]
A local model that can do better than Siri or Alexa as a personal or home assistant is, in my eyes, very useful. Being able to run on a phone or watch or glasses translates to me, low-powered AI, and not necessarily that I want my phone, or watch, or glasses to run things for me.
My Siri use has narrowed down to just setting timers. And even then, I still have my phone call people in the middle of the night. Siri is pretty dumb and does not do what I want it. I’d rather be able to customize an assistant to myself.
I am also thinking of automation in my day to day workflow for work.
jcgrillo 1 hours ago [-]
OK.. but what would you have all this "automation" actually do? What is Siri failing to do that you want it to do? How would customizing an assistant (for whatever definition) help?
jasonjmcghee 4 hours ago [-]
Throwing a few things out - HN has changed over the years, but people make stuff to make stuff. There don't need to be product use cases. The tone of the comment goes against the spirit of HN - likely the reason for downvotes.
That aside- a very small model that takes text and outputs structured json according to a spec is nice. It let's you turn natural language into a user action. For example, command palettes could benefit from this.
If you can do a tiny bit of planning (todo) and chain actions, it seems reasonable that you could traverse a rich state space to achieve some goal on behalf of a user.
Games could use something like it for free form dialog while stool enforcing predefined narrative graphs etc.
I'm sure you could come up with more. It's a fuzzy function.
jcgrillo 4 hours ago [-]
> people make stuff to make stuff. There don't need to be product use cases.
OK. Great! So it doesn't need to be a commercial product. But does it do something (anything?) interesting? I'm interested in your games example, I'd love to see it done in real life. IIUC, game AIs are actually much more constrained and predictable for play-ability reasons. If you let it go all free form a plurality of players have a "WTF??!?" experience which is super Not Good.
digdugdirk 3 hours ago [-]
It doesn't have to do any thing interesting - it's completely fascinating all on it's own. If you understand anything about the math and science behind LLMs, you'll understand that this is an achievement worthy of sharing to a community like HN.
That being said, small models like these have plenty of use cases. They allow for extra "slack" to be introduced into a programmatic workflow in a compute constrained environment. Something like this could help enable the "ever present" phone assistant, without scraping all your personal data and sending it off to Google/OpenAI/etc. Imagine if keywords in a chat would then trigger searches on your local data to bring up relevant notes/emails/documents into a cache, and then this cache directly powers your autocomplete (or just a sidebar that pops up with the most relevant information). Having flexible function calling in that loop is key for fault tolerance and adaptability to new content and contexts.
Its cool. Enjoy it.
jcgrillo 3 hours ago [-]
> Something like this could help enable the "ever present" phone assistant, without scraping all your personal data and sending it off to Google/OpenAI/etc
OK so show me what that's for. Show me something useful you can do with that ability.
> Imagine if keywords in a chat would then trigger searches on your local data to bring up relevant notes/emails/documents into a cache, and then this cache directly powers your autocomplete (or just a sidebar that pops up with the most relevant information).
I'm really trying but.. idgi? I truly cannot imagine how this would improve my life in any way...
> Its cool. Enjoy it.
No. It sounds like a useless complication on my watch. I don't fucking care if it can tell me the phase of the moon. I can look up at the sky and see the moon and know what phase it is.
EDIT: You say:
> If you understand anything about the math and science behind LLMs, you'll understand that this is an achievement worthy of sharing to a community like HN.
OK. So educate me. Tell me what I'm missing.
HenryNdubuaku 6 hours ago [-]
You can think of “phone use” for instance, what Siri is supposed to be.
jcgrillo 6 hours ago [-]
I mean.. Siri basically works? When I'm driving I say "Hey Siri, find me a gas station along my route", and it does. Or I say "Hey Siri, call Joe Bob mobile" and it does. Or I say "Hey Siri, play me a podcast". This is kind of a solved problem already? When I'm driving this is literally as complicated of a distraction as I want--I'm not going to be dictating emails or texts. When I'm not driving, the touchscreen keyboard (as shitty an interface as that is) is 100x better than voiced natural language commands.
ilaksh 6 hours ago [-]
It does just barely work now after they spent billions, and they may still fall back to cloud LLMs for a significant number of things. This is a way that everyone can get that on the actual Apple Watch or local phone for any application they build.
jcgrillo 5 hours ago [-]
I get that, but I still can't imagine what it might be for. TBH I don't have a smart watch, because I can't think of anything I'd want one for--my mechanical watch keeps time to within a few seconds per month and the lume lasts all night. I don't know what making it "smarter" would do for me, it does an A+ job of being a watch. What are the things that "everyone" can build with this that actually matter? Like, what is the differentiator?
EDIT: To be clear, the monoculture of phone operating systems sucks. If this somehow enables more entrants into that space then I'm all for it. However, I don't see this in particular being the deciding factor... For example, the reason I don't run a 3rd party operating system on my phone isn't because it's lacking Siri or "OK Google" (if these things went away tomorrow I'd barely notice), it's because it would be a pain in the ass to make it be a phone.
logdahl 9 hours ago [-]
I find this stuff super fascinating and been thinking about it myself. Maybe one could bootstrap tiny models on a rather 'pure' procedural data set. Neglecting [0] of course...
I'd expect either a chain load or just a 2 hour timer. Further attempts humorously give two separate 1-hour-timers.
8 hours ago [-]
syntaxing 7 hours ago [-]
This would be amazing for home assistant.
synesthesiam 6 hours ago [-]
On my list to check out tomorrow :D
syntaxing 4 hours ago [-]
Wow can’t believe the voice engineer lead for Nabu Casa is here! Super excited to see if this works for HA!
HenryNdubuaku 6 hours ago [-]
Thanks, keep me posted!
sroussey 5 hours ago [-]
Can this be converted to onnx or otherwise be used in a browser?
quadrature 8 hours ago [-]
Does the model have capacity for in context learning ?, if we give it examples of patterns can it follow them ?.
HenryNdubuaku 8 hours ago [-]
Not yet, for now. But it’s in the works!
dangoodmanUT 7 hours ago [-]
Why pick Gemini? It's probably the worst tool calling model of the major labs.
HenryNdubuaku 7 hours ago [-]
Cheaper APIs
halyconWays 2 hours ago [-]
I assume this would only be useful as the second stage after a model like Whisper, as it can't understand speech where you'd want it, like on a phone or small device?
roggenbuck 7 hours ago [-]
This is some excellent work Henry! Very excited to try it out.
HenryNdubuaku 7 hours ago [-]
Thanks, let me know how it goes!
ac29 9 hours ago [-]
FYI, distilling Gemini is explicitly against the ToS:
"You may not use the Services to develop models that compete with the Services (e.g., Gemini API or Google AI Studio). You also may not attempt to reverse engineer, extract or replicate any component of the Services, including the underlying data or models (e.g., parameter weights)."
Havoc 9 hours ago [-]
Yeah I think Google should shove that somewhere. They effectively distilled all the internet's knowledge into these models...without asking & without permission
HenryNdubuaku 9 hours ago [-]
Thanks, Needle doesn’t compete with those tools though and the distillation process did not access the weights.
ilaksh 9 hours ago [-]
I think GLM 5.1 or Kimi 2.6 could substitute for this type of purpose.
iAMkenough 8 hours ago [-]
FYI, Gemini was developed using stolen copyrighted works without author consent. The double standard is striking.
ForHackernews 9 hours ago [-]
So is copying all the books in the world.
xgulfie 9 hours ago [-]
This is being downvoted but it's worth noting if only for the "be careful" aspect.
That said, we need more people distilling models IMO, just be ready for a C&D and a ban
vablings 9 hours ago [-]
Oh no! They stole the model weights!
Distillation "attacks" is such bullshit
Man, I love that there are still people writing new MOO servers in 2026. Any game out there already running on mooR?
cmrdporcupine 8 hours ago [-]
Many people tease that they will, and start... but then kinda stop. But mostly just been building my own bespoke thing on my own bespoke platform, and kinda running out of steam because I need to make $$ instead.
HenryNdubuaku 9 hours ago [-]
Thanks, let us know how it goes!
deepsquirrelnet 9 hours ago [-]
This is really cool. Any plans to release the dataset?
HenryNdubuaku 9 hours ago [-]
We include the dataset pipeline in the codebase so far, might release dataset.
theykk 6 hours ago [-]
hey nice work, is it possible to release the datasets?
HenryNdubuaku 5 hours ago [-]
We have so far released the dataset generation code
varispeed 7 hours ago [-]
What is the use case for this?
masafej536 1 hours ago [-]
Something like this together with MCP can replace APIs for 3rd party integrations.
You just give it instructions to "post a message in slack" and provide it slack MCP tools and it figures out the rest on its own. No need to read up on slack API docs or worry about breaking changes.
HenryNdubuaku 7 hours ago [-]
Deploying AI on tiny devices like watches, earphones, glasses etc.
varispeed 6 hours ago [-]
Ok, but why? What is the use case?
chris_money202 6 hours ago [-]
I don't think the limit is just on tiny devices. It can also be used in apps on generic computers, because its so small anything can run it reasonably quick.
For example, I am thinking this could be helpful for say if you have a complicated build and test infrastructure, fine tune this model on that infrastructure and then people can say more generic things like build and run this library's test, rather than issuing the exact commands to do that or going to Claude, GHCP, etc
BoredPositron 8 hours ago [-]
I source old, defective high-end radios with timeless designs from brands like Grundig or Braun, and replace the original hardware with a Raspberry Pi while using the original audio parts to build custom smart speakers. Reliable hotword detection and voice command recognition have been a persistent challenge over the years, but whisper and other small models have helped enormously. At the moment I have ollama running on my server with qwen 9b which works fine but a 26M that could be deployed on the pi itself would be amazing.
HenryNdubuaku 7 hours ago [-]
Sounds cool, play with it and let uk know what you think!
The examples are things like "What is the weather in San Francisco", where you are only passed a tool like
I had a thing[1] over 10 years ago that could handle this kind of problem using SPARQL and knowledge graphs.My question is how effective is it at handling ambiguity.
Can I send it something like a text message "lets catch up at coffee tomorrow 10:00" and a command like "save this" and have it choose a "add appointment" action from hundreds (or even tens) of possible tools?
[1] https://github.com/nlothian/Acuitra/wiki/About
input: i need to contact my boss i will be late. output: [{"name":"send_email","arguments":{"to":"boss@company.com","subject":"Running late","body":"I will be late for the meeting."}}]
it did have the send_email tool on the left hand side though
Google reportedly reacts to distillation attempts "with real-time proactive defenses that can degrade student model performance". So if they detected you, they could have intentionally fed you a dumber but plausible variant of Gemini: https://cloud.google.com/blog/topics/threat-intelligence/dis...
But also, this model is small and just focusing on the tool use. In terms of token usage, you're probably not anywhere near the people that are trying to distill the entire model.
Sonnet would often call tools quickly to gather more context, whereas Opus would spend more time reasoning and trying to solve a problem with the context it had.
This led to lots of duplicated functions and slower development, though the new models (GPT-5.5 and Opus 4.6) seem to suffer from this less.
My takeaway was that “dumber” (i.e. smaller) models might be better as an agentic harness, or at least feasibly cheaper/faster to run for a large swath of problems.
I haven’t found Gemini to be particularly good at long horizon tool calling though. It might be interesting to distill traces from real Codex or Claude code sessions, where there’s long chains of tool calls between each user query.
Personally, I’d love a slightly larger model that runs easily on an e.g. 32GB M2 MBP, but with tool calling RL as the primary focus.
Some of the open weight models are getting close (Kimi, Qwen), but the quantization required to fit them on smaller machines seems to drop performance substantially.
But it's really interesting to me that that may be possible now. You can include a fine-tuned model that understands how to use your program.
E.g. `> toolcli what can you do` runs `toolcli --help summary`, `toolcli add tom to teamfutz group` = `toolcli --gadd teamfutz tom`
Today is 2026 after all
You can check the very simple docker file there.
https://news.ycombinator.com/newsguidelines.html
As the other poster noted, the post wasn’t meant to be read as a personal attack
Why are you attacking me?
I personally prefer the M to the B. I guess as an engineer, noticing the units comes pretty naturally.
Announcing something that's 1/1000th is significant and remarkable! Hiding it in a single letter is burying the lede.
I have been building for small (20B or less) models for quite a while. Highly focused/constrained agents, many of them running together in some kind of task orchestration mode to achieve what feels like one "agent".
I build (privacy first) desktop apps this way and I want to get into mobile apps with similar ideas but tiny models.
Heh, what a coincidence, just today one of my students presented research results which also confirmed this. He removed MLP from Qwen and the model still could do transformation tasks on input but lost knowledge.
What is a distilled model?
Why doesn't Google do this (to make their models smaller)?
Seems like you could make a competitor to Gemini?
1. Distilled means taking the intelligence of a big model and compacting into a tiny model.
2. Google already does so with FunctionGemma, but Needle argues that better performance could be achieved with 10x smaller model using our technologies.
Smaller model requires less space on disk, less video memory, and less compute (cheaper hardware).
Downside is that distilled model performs worse on the same benchmarks compared to original model.
> Repository Not Found for url: http s://huggingface.co/api/datasets/Cactus-Compute/needle-tokenizer/revision/main.
Got a bunch of errors trying to run it on CPU though. Very likely connected to me running this in a container (unpriv LXC), but figured for 26M CPU would suffice.
https://pastebin.com/PYZJKTNk
I haven't played with it yet, but does it ever return anything other than a tool call? What are the failure modes? What if it doesn't understand the request? Does it ever say it can't find a tool? Does it get confused if there are two similar (but different) tools? Can it chain tools together (e.g. one tool to look up and address and another to get directions to the address)?
I mean, I plan on downloading the model later tonight and finding out for myself, but since I'm stuck at work right now, I figured I'd ask anyway...
Come to think of it, this could be a nice model to have as the first pass in a more complex agent system where Needle hands of the results of a tool call to a larger model.
I will defiantly play around with this!
Are you Calvin or Hobbes?
But how do you get 'Paris' into the value vector in that case? The value vector is just the result of a matrix multiplication, and without a nonlinearity it can't perform a data-dependent transformation. Attention still acts as a nonlinear mixer of previous values, but your new output is still limited to the convex combination of previous values.
Ok wait I think I see what you mean. Although maybe it's not getting paris _into_ the value vector that's hard, but isolating the residual stream to _only_ that instead of things like other capitals.
So as a naive example maybe at the very first layer consuming your tokens: Q{France} would have high inner product with K{capital} and so our residual would now mostly contain V{capital}, which maybe contains embeddings of all the capitals of all countries. You need some way to filter out all the other stuff, but can't do that without a FFN + activation.
Just throwing in a relu by itself won't help since that would still work on all the elements uniformly, you need some way to put weight on "paris" while suppressing the others, i.e. mixing within the residual stream itself.
Although maybe if you really stretch it, somewhere in a deeper layer you could have 1-hot encoded values with a "gain" coefficient so that when you do the residual addition it's something like {<paris>, <tokyo>, <dc>} + 10000*{<1>, <0>, <0>} and then if you softmax that you get something with most of its mass on "Paris". But it seems like this would not be practical, or it's just shifting the issue to how that the right 1-hot vector is chosen
Is it a replacement for Kimi 2.7, Claude Haiku, Gemini Flash 3.1 lite, a conversational LLM for the situations where it's mostly tool-calling like coding and conversational AI?
EDIT: more of you cretins have downvoted than have replied.. so.. show your cards.
My Siri use has narrowed down to just setting timers. And even then, I still have my phone call people in the middle of the night. Siri is pretty dumb and does not do what I want it. I’d rather be able to customize an assistant to myself.
I am also thinking of automation in my day to day workflow for work.
That aside- a very small model that takes text and outputs structured json according to a spec is nice. It let's you turn natural language into a user action. For example, command palettes could benefit from this.
If you can do a tiny bit of planning (todo) and chain actions, it seems reasonable that you could traverse a rich state space to achieve some goal on behalf of a user.
Games could use something like it for free form dialog while stool enforcing predefined narrative graphs etc.
I'm sure you could come up with more. It's a fuzzy function.
OK. Great! So it doesn't need to be a commercial product. But does it do something (anything?) interesting? I'm interested in your games example, I'd love to see it done in real life. IIUC, game AIs are actually much more constrained and predictable for play-ability reasons. If you let it go all free form a plurality of players have a "WTF??!?" experience which is super Not Good.
That being said, small models like these have plenty of use cases. They allow for extra "slack" to be introduced into a programmatic workflow in a compute constrained environment. Something like this could help enable the "ever present" phone assistant, without scraping all your personal data and sending it off to Google/OpenAI/etc. Imagine if keywords in a chat would then trigger searches on your local data to bring up relevant notes/emails/documents into a cache, and then this cache directly powers your autocomplete (or just a sidebar that pops up with the most relevant information). Having flexible function calling in that loop is key for fault tolerance and adaptability to new content and contexts.
Its cool. Enjoy it.
OK so show me what that's for. Show me something useful you can do with that ability.
> Imagine if keywords in a chat would then trigger searches on your local data to bring up relevant notes/emails/documents into a cache, and then this cache directly powers your autocomplete (or just a sidebar that pops up with the most relevant information).
I'm really trying but.. idgi? I truly cannot imagine how this would improve my life in any way...
> Its cool. Enjoy it.
No. It sounds like a useless complication on my watch. I don't fucking care if it can tell me the phase of the moon. I can look up at the sky and see the moon and know what phase it is.
EDIT: You say:
> If you understand anything about the math and science behind LLMs, you'll understand that this is an achievement worthy of sharing to a community like HN.
OK. So educate me. Tell me what I'm missing.
EDIT: To be clear, the monoculture of phone operating systems sucks. If this somehow enables more entrants into that space then I'm all for it. However, I don't see this in particular being the deciding factor... For example, the reason I don't run a 3rd party operating system on my phone isn't because it's lacking Siri or "OK Google" (if these things went away tomorrow I'd barely notice), it's because it would be a pain in the ass to make it be a phone.
[0]: http://www.incompleteideas.net/IncIdeas/BitterLesson.html
Result: [{"name":"set_timer","arguments":{"time_human":"1 hour"}}]
Query: in 1 hour set a timer for 1 hour
Result: [{"name":"set_timer","arguments":{"time_human":"1 hour"}}]
I'd expect either a chain load or just a 2 hour timer. Further attempts humorously give two separate 1-hour-timers.
"You may not use the Services to develop models that compete with the Services (e.g., Gemini API or Google AI Studio). You also may not attempt to reverse engineer, extract or replicate any component of the Services, including the underlying data or models (e.g., parameter weights)."
That said, we need more people distilling models IMO, just be ready for a C&D and a ban
For example, I am thinking this could be helpful for say if you have a complicated build and test infrastructure, fine tune this model on that infrastructure and then people can say more generic things like build and run this library's test, rather than issuing the exact commands to do that or going to Claude, GHCP, etc