Top
Best
New

Posted by HenryNdubuaku 5 days ago

Show HN: Cactus Needle 3: 8-29MB automation models can match DeepSeek V4 Flash(cactuscompute.com)
Hey HN, Henry from Cactus here.

We submitted Needle 2 here a few weeks ago, and the feedback in the discussion thread was incredibly valuable, thanks! Thanks to all that feedback, we’ve been able to move quickly to release Needle 3 and I'd love to hear what you think again.

The key features:

1) Automation (tool calls & structured JSON output): Needle still doesn't chat by design, its quite challenging to pack general capacity into such small models, so we focus on tool calls and structured JSON. If no tool you declared fits the request, you get an empty list back (note for when playing with the demo).

2) Intelligence Laddering: Every layer (2 to 20) is a deployable subnetwork, so one set of weights, 25 to 121 million parameters at 2-bit, shipping as 8-29MB binaries. On a Raspberry Pi 5 it decodes at up to 4k tokens/sec and prefills at up to 10k.

3) Monarch Hadamard MLP: replaces the dense FFN with three learnable Walsh-Hadamard-initialized Kronecker (Monarch) factor pairs interleaved with per-channel diagonal scales, fixed permutations, a SiLU nonlinearity, and a rank-8 input-conditioned gate, so each token gets a fully mixed nonlinear transform of its d_model channels at O(d√d) parameters and compute instead of the O(d²) a dense 4x-expansion MLP would cost.

4) Performance: On Mobile Actions (phone commands, scored on the exact call) the 20-layer model gets 86.0 through the shipped 2-bit binary; LFM2.5 1.2B is at 82.4, Qwen3.5 0.8B at 76.0, Apple's on-device model at 57.6, all at f16. More results on the link, we do not win everywhere ofc.

5) Multilingual: Needle 3 now supports English, French, Spanish, German, Dutch, Italian, Polish, with more languages coming.

6) Finetuning: You can achieve DeepSeek v4 Flash grade performance on a narrow task with just 4L, stress on "narrow task", we found that production users often prefer tuning before production.

7) Triggers: Grounding is a common challenge for tool call, at least for Needle 2, so we added support case-insensitive regular expressions matched against each request to gate false negatives.

8) Confidence: Every response also carries a calibrated confidence score, the minimum of a judgement on the finished call and its decode probability. Act above your threshold, show the call and ask below it, or escalate to a bigger model.

9) Supported Platforms: macOS, Linux on x86-64, ARM64, ARMv7, RISC-V and MIPS32, Windows x64 and ARM, Android, iOS, watchOS, tvOS, the browser as WebAssembly, and a WASI component.

Thanks for reading and as always, thoughts appreciated!

236 points | 92 comments
IanCal 4 days ago|
Wondered if it'd turn on the lights in the bathroom with these:

"I need a wee" -> tries to play music because "wee" is a genre

"I need a wee wee" -> starts the vaccuum in the bathroom

"I'm going to the toilet" -> says it'll turn on the toilet, and I'm not totally sure what that entails.

"I'm going to the toilet and can't see" -> reasons that lights should be on in the bathroom, then chooses again to turn on the toilet.

"I'm going to the toilet and can't see where I'm going" -> reasoning is "'going to the toilet' -> control_device with device 'coffee maker' (toilet implies coffee maker)"

"I'm going to the toilet and can't see where I'm going because it is too dark" -> "'dark' -> direction 'dark'; adjust_lights with brightness 100 for darker light"" and chooses to turn the lights in the living room to "dark" which fails.

At this point the vacuum is in a dark bathroom, the living room is 100% brightness and playing "wee". At least there's coffee.

ash_091 4 days ago||
Pretty much matches my experience.

> 'sleepy time' means sleeping → start_vacuum with room 'bedroom' to start cleaning

The "DeepSeek 4 Flash grade" claim seems far fetched.

HenryNdubuaku 4 days ago|||
thanks for these haha, you can actually edit the tools and/or their descriptions, the demo is just a "get started" preset. But still we do have room for reasoning improvement!
IanCal 4 days ago||
What kinds of things do you expect to work?

Edit - I’m struggling to get anything useful. Reasoning is often utter nonsense and the actions are very often very wrong. To the point of seemingly needing very precise sentences to work at which point you may as well do regexes. Very simple things like clean one room then another with the vac fails.

HenryNdubuaku 4 days ago||
Thanks for the feedback! Implications and relations are hard for the model to understand (things like go to the living room, then the kitchen, and back), so yes the cleanest use cases involve direct language. Reasoning isn't true reasoning in the way general LLMs do it, it is more like grounding for the model that it generates itself. This can often become nonsensical specifically when the model gets things wrong, providing signal to the confidence.
rohansood15 4 days ago||
Can you share an actual example of where it works please?
electroglyph 4 days ago|||
those are all expecting far too much for models this size
p1necone 4 days ago||
Given the title of the post says 'can match deepseek v4 flash' I think it's fair to call out these sort of dumb mistakes.
mentalgear 4 days ago||
Maybe with fine tuning?
gs17 4 days ago||
"turn all the lights on/off" and "it's too dark in the bathroom" worked for me, but anything less direct didn't. "it's too cold" actually made it turn the thermostat down ("it's cold" made it... turn the lights down?)! Although the confidence on the bad responses was pretty low, so it might be worth adding a threshold to the demo.

Or maybe it just has a weird thermostat down bias? "make it hot" also had it turn it down (specifically it went from 20->18, or at least tried to, the UI still showed 20), with high confidence. Also might have a bit of a Celsius vs Fahrenheit confusion. Neat concept, but I might not want to let it control the oven at the moment.

The laptop demo worked better until I tried to open the mail app. "Check mail" kept opening the browser with an error, and "check email" makes a note with the text "email", "open email" goes to "https://api.email.com/v1/email" in my real browser, but "open mail" does work.

And I presume the "reasoning" isn't very trustworthy? In the car I got "'turn it up' means lower volume -> set_volume with lower value." For the house, reasoning would correctly say that I wanted the alarm off, but it didn't actually do it.

HenryNdubuaku 4 days ago|
Hey, thanks a lot for this feedback, very useful and actionable for us! Quite a few of these came down to our tool definitions in the playground as well as out triggers. We updated them just now and these should be more reliable. Really this goes to show that needle shines through after putting in the work to make the tool list around it good for your use case. As for the reasoning, yes its main function is really to provide more words/keywords that the model can latch onto when generating the tool call response, since this is a SAN model it needs more grounding in existing context.
IanCal 4 days ago||
I can’t help but wonder how well more traditional approaches would do with this. Something like a map of statements to actions, with fuzzy search - then remove what used to be the labour intensive part of this by handing it to a decent llm to generate the sentences.
HenryNdubuaku 4 days ago||
That's a really good point and I think it's not yet clear how well, say, 8-30MB worth of regexs with accompanying algorithmic structure would do on these tasks. I would imagine they do quite well on a well defined task, but it would be much harder to then adapt this set to a new domain. A big part of Needle's promise is how easy it is to finetune. Ultimately I think the two approaches can be more complimentary to each other, rather than choosing only one (see triggers!).
potatoman22 4 days ago||
I think a good "traditional" approach would look like a BM25 algorithm over an index of trigger phrases for each category, sitting behind a majority-vote classifier. The "fine tuning" would be done by reindexing the data, generating different/new phrases, and tuning the classification threshold.
HenryNdubuaku 4 days ago||
I think we might look into creating a baseline like this for our future models
raybb 4 days ago||
I have an idea for a use case for this, and I'm wondering if you think it makes sense or if you have any thoughts on the approach.

I'm a big fan of OpenStreetMap, and I enjoy editing it from my computer. From my phone, I find it quite tedious trying to make sure I type in the phone number exactly correctly and double-check it, or find and select the right field from the large list of fields available in Upredor.

Generally, how it works is I see a restaurant, and there's a sign. I know that it says, "Cash only. Here's the phone number. Here's the opening hours." What would be really cool is if I could just speak to the phone and say, "Hey, here's the information about this place." It would automatically use your location to detect what places are nearby and maybe even detect which place you're talking about, and then tell you, "Okay, here are the changes I think you're proposing to make, or these things you stated are the ones that would create a diff." This would be limited to just perhaps the 20 most common keys in some predefined set of values for most of them. Like cuisine=x should just match to the most common not make up new ones.

Of course, this is something a large language model could do, but having it run on device would be a lot nicer and cheaper.

HenryNdubuaku 4 days ago|
Makes a lot of sense! declare one record with the ~20 keys as fields, cuisine and friends as enums with the common values, and the grammar can't produce a value outside the set; fields with no evidence come back empty, so the output is exactly the diff.

For "which place", query nearby POIs from location in the app and pass the candidate names as an enum field, so Needle picks rather than guesses.

Two caveats: it's text-in, so you need on-device STT first, and opening_hours syntax is the risky bit, so either put the format in the description or capture the raw hours and normalise in code.

hirako2000 4 days ago||
My thought, the growing number of dubious claims that a tiny model beats LLMs will make any useful innovation be overlooked.

What's more important than the resource requirements is to highlight what the model simply cannot even attempt to do that general LLMs do decently well.

In other words, tell me the anti use case clearly so that I don't have to find out myself.

janalsncm 4 days ago||
An LLM is a Swiss Army knife. This is a corkscrew.

All of the other tasks a general-purpose LLM can do (write me a poem about pizza, rewrite this code in rust, tell me about the causes of the war of the roses) are unsupported.

The only use case this supports is converting unstructured text into structured json calls, and doing that quickly in a low memory environment.

HenryNdubuaku 4 days ago||
Strong point! Needle is a task-specific model and bullet 6 stressed that it is only trained to be good on a set of narrow tasks, but I guess it could be clearer?
owebmaster 4 days ago||
It would be clearer if you didn't use AI to reply.
Scaevolus 4 days ago||
This is a solid improvement over Needle 2, which I tried using for a tool-calling interface to a Runescape database site. Unfortunately it's still not quite capable enough for my target compared to FunctionGemma.

   Model                            Correct tool shape    Exact arguments
  ━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━  ━━━━━━━━━━━━━━━━━━━━  ━━━━━━━━━━━━━━━━━
   FunctionGemma fine-tune, BF16       209/230 (90.9%)    196/230 (85.2%)
  ───────────────────────────────  ────────────────────  ─────────────────
   Needle 3 fine-tuned W4A8             74/230 (32.2%)     47/230 (20.4%)
  ───────────────────────────────  ────────────────────  ─────────────────
   Needle 2 fine-tuned W4               59/230 (25.7%)     43/230 (18.7%)
HenryNdubuaku 4 days ago|
Thanks for testing Needle out! I'd be very interested in hearing more about the finetuning setup to see how we can make both the library's finetuning setup and the model better.
neilellis 4 days ago||
I tried this today for labelling - and for that task it was very bad MNLI was better - so you are going to need to match the use case for this pretty exactly. (at 29MB params one would expect that!) I'm obviously not saying labelling is a good use case :-) just adding a data point.

Jev has put the cat amongst the pigeons so suddenly everyone is looking at classifiers and encoder only models again.

My ideal model would be a general purpose LLM API that can answer classification questions and as it does so distils to an encoder only model so that the more classifications I do the cheaper it gets (i.e. the more it offloads to the classifier). If anyone ever wants to do this as a service do let me know, because it's just another piece of code to manage in each new project that needs classification.

Also a model that could do this internally would be nice :-)

HenryNdubuaku 4 days ago|
Hey! Yeah I think for labelling the model would need to have much better world knowledge than its current size allows. Jev really is a very good model, I think it has a very strong place in the upcoming tech stacks. Really good suggestion to make a continuously distilled model, we are going to have to look into that one :)
neilellis 4 days ago||
Good luck with this model/product, in the excitement of LLMs people seem to forget applicability. I very much like to see innovation in this space, so well done!
Retro_Dev 4 days ago||
A very cool project, but of course not perfect. I'd rather have 30 megabytes of phrases mapped to the perfect and correct control changes in a home, rather than a heuristic built around 30 megabytes. I tried to "warm the house" (increase the temperature of the thermostat), but the model actually turned the lights to a "warm brightness" - reasoninig being `"'warm the house' -> set_lights to warm brightness. No specific room given, so use default 'living room' as default."`
HenryNdubuaku 4 days ago|
Thank you fr this. "warm the house" now goes to the thermostat. It's fair that a more deterministic system with just action phrases would be easier to debug/interpret, but I think there is room for both a model that is trained to understand meaning as well as deterministic logic aiding it. To this end, we just started exploring the idea of triggers, and are working towards expanding this even more.
janalsncm 4 days ago||
Hey, I’m really happy that someone is building this. I tried doing something similar a couple of months ago and came to the conclusion that the dataset was at least as important as the modeling itself. Building a good dataset is nowhere near as flashy as building a novel model architecture, but it really is critical.

For instance, you want to be able to handle any smart home commands people could issue, right? What are all of the smart home devices? What are all of the ways people might want to issue commands? Also, for things like Spotify, it’s not going to know what “The Beatles” are or “Led Zeppelin”. Artist and song names themselves are easily just as hard as all of the smart home devices combined.

The simple attention network stuff is cool, it makes sense to drop the MLP when it dominates the param count. But you’ll definitely lose some “world knowledge”. That’s probably ok though.

HenryNdubuaku 4 days ago|
100%, data was honestly most of the work, Needle 3 is trained on 360B tokens of structured data and we spend way more time on the generation pipeline than on the model. On Led Zeppelin, Needle doesn't actually need to know it, arguments are copied from the request so it just lifts the name into the artist field. The knowledge went into the engram btw, 70M of the 121M params are n-gram tables, so it can tell artist vs song without an MLP. Also yes, "play their second album" won't work, that needs the world knowledge it doesn't have.
viccis 4 days ago||
I'll try to get something set up to try this out. I've been working on an ESP32 based Echo replacement that sends audio back to a backend server I run, and one question I had was whether models small enough to run on a Mac Mini or even smaller hardware are good enough to handle basic tool calling functionality with a bit of reasoning where needed.

I have a test suite that tries like ~36 different scenarios, including things like starting multiple timers, saying "actually cancel that timer" and whether it knows to do that one you just created. Basic decision making on top of tool calling. I found so far that, for example, Qwen3.8 on my local machine does pretty poorly even relative to Gemma4 E4B (~9.6gb) and that the best price/performance outcome I've found so far with openrouter is actually GPT Luna, but obviously I'd love to get something that works as well running locally for privacy reasons.

Would love to try this out, I'll just need to tweak my benchmarker to use however this serves it.

HenryNdubuaku 4 days ago||
Hey there! If you end up trying out needle on the test suite it would be very useful for us if you could share some failure modes of the model! We are always trying to understand where the model isn't doing good and where we can make it better.

For your question on tool calling, I think you will find that the model is pretty good at simpler tool calls and parallel ones, but can struggle with implied references and multistep reasoning. These are definitely things that can improve with task-specific finetuning but for some things you just have to have a model that is properly sized. That said, we are always trying to improve the model so that it can handle an ever larger set of queries

viccis 4 days ago||
Thanks for responding. I ran it just now and it looks like it could be useful if I change and limit the scope of the kinds of actions I need. Here is a breakdown of where it struggled vs some local models (~8-12B parameters running on a 16GB Mac Mini or my desktop's 3080), bear in mind I had Claude integrate it into the benchmarks and this is its interpretation, not my own:

What it gets wrong:

- It copies numbers instead of converting them. "25 minute timer" becomes duration_seconds: 25, and "twelve minutes" becomes 120. The first one comes with 100% confidence.

- It picks the wrong action. "take the paper towels off the list" became an add. "remind me in 20 minutes" became a timer. "add five minutes to the pasta timer" became a new timer plus a cancel.

- It never declined anything with our full tool set. Background chatter became note_save "blue one" at 0.99 confidence. "play some jazz" became a screen card, and "wake me up at 6 30" became a 630-second timer.

- It can't use household context. Notes, timer names and reminder IDs have no place in its input. Passing them anyway made results worse (5 of 27 single-turn requests right, versus 8 of 27 without), so the backend now leaves them out.

- Follow-ups mostly broke. "take off the last one" removed the whole list.

HenryNdubuaku 4 days ago||
This is extremely useful feedback for us, thanks! I think the easiest thing here that can be fixed with tool definitions is the number conversions. Additionally, the model tends to work better with fewer tools. We will definitely be focusing on better context usage and followups going forward as well.
akadeb 1 day ago||
hey viccis the tool calling on ESP32 with a Mac M4 is exactly what I built here. check it out and lmk if it helps https://github.com/akdeb/open-toys
lostmsu 3 days ago|
> Monarch Hadamard MLP: replaces the dense FFN with three learnable Walsh-Hadamard-initialized Kronecker (Monarch) factor pairs interleaved with per-channel diagonal scales, fixed permutations, a SiLU nonlinearity, and a rank-8 input-conditioned gate, so each token gets a fully mixed nonlinear transform of its d_model channels at O(d√d) parameters and compute instead of the O(d²) a dense 4x-expansion MLP would cost.

Wow, I was just researching W-H in transformers. Did yours seem to work? In my experiments swapping various components for W-H-like transforms caused extreme quality degradation.

UPD. according to the comments here, this model simply does not work at all, so I guess the answer is NO

More comments...