Cactus releases Needle 3, 8-29MB models for on-device tool calling
Tiny automation model trades general chat for tool-calling performance matching larger systems, targeting mobile and embedded devices.
What to know
- Cactus Needle 3 is an 8-29 MB model specialized exclusively for tool calling and structured extraction on mobile and embedded devices, sacrificing general chat capability for task performance.
- Intelligence laddering allows developers to choose layer depth (2-20) with one set of weights; the 4-layer version claims to match DeepSeek V4 Flash when fine-tuned.
- Early demo testing reveals consistent tool-routing failures on natural language commands, with users reporting the model misinterprets intent (e.g., "warm the house" → lights, "I need a wee" → music).
- Community skepticism emerged over the DeepSeek V4 Flash performance claim, with users and developers questioning whether small model heuristics can reliably replace rule-based automation systems.
The dispute Whether Needle 3's claimed performance parity with DeepSeek V4 Flash is credible given observed tool-routing failures in the public demo. · positions read across 20 posts and comments
The tool-routing failures in demo testing undermine claims of DeepSeek V4 Flash parity and suggest rule-based systems are more reliable.
-
“The "DeepSeek 4 Flash grade" claim seems far fetched.”
ash_091 · Hacker News ↗
Needle 3 has real potential for on-device voice and automation in embedded systems despite current reasoning limitations.
-
“Esp paired with a small Whisper or Parakeet voice model, this enables some amazing real-world use cases in lower-power situations (car, marine, home, PLC, industrial automation).”
jamiesonbecker · Hacker News ↗
Dataset quality and task-specific fine-tuning matter as much as model architecture for reliable tool calling.
-
“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.”
janalsncm · Hacker News ↗
Cactus Compute AI model developerHenry Ndubuaku Founder, CactusDeepSeek Competing model vendor (V4 Flash)
How it unfolded 7 developments, newest first · click a bar or a number to jump articlespostscomments
-
7
Positive feedback on use case potential for embedded voice control
Despite mixed demo results, a developer highlighted practical potential when paired with small voice models for on-device automation in cars, homes, marine, and industrial systems.
“Esp paired with a small Whisper or Parakeet voice model, this enables some amazing real-world use cases in lower-power situations (car, marine, home, PLC, industrial automation).”
— jamiesonbecker -
This is incredibly cool. I tried telling it to both turn up the temperature in the home and turn down the lights in the LR and got a beautiful JSON doc.Esp paired with a small Whisper or Parakeet voice model, this enables some amazing real-world use cases in lower-power situations (car, marine, home, PLC, industrial automation). Combining with a…
2 more of the top 3 · 6 posts in this stretch
-
site:apps.dos.ny.gov "BUFFALO" "CORPORATION" apps.dos.ny.gov<meta http-equiv="refresh" content="0; url=
-
"reasoning": "'burn us all alive' means turn off all lights -> set_lights with room 'all' and state 'off'."
-
-
6
User prefers rule-based mappings to heuristic model for home control
A developer noted preference for deterministic phrase-to-action mappings over the model's heuristic approach, citing an example where "warm the house" was misinterpreted as a lighting brightness command rather than a thermostat adjustment.
“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.”
— Retro_Dev -
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…
2 more of the top 3 · 3 posts in this stretch
-
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" -…
-
"Illuminate (roomname), de-illuminate (roomname)" works well. Harden perimeter, locks doors and sets alarm.Neat stuff.
-
-
5
Skepticism over "DeepSeek V4 Flash grade" claim amid demo failures
As more users tested the demo and encountered routing errors, skepticism grew about the claimed performance parity with DeepSeek V4 Flash. One commenter noted the claim "seems far fetched" given observed failures.
“The "DeepSeek 4 Flash grade" claim seems far fetched.”
— ash_091 -
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…
2 more of the top 3 · 4 posts in this stretch
-
Pretty much matches my experience.> 'sleepy time' means sleeping → start_vacuum with room 'bedroom' to start cleaningThe "DeepSeek 4 Flash grade" claim seems far fetched.
-
What is 8 dash 29 MB? And the copy on the landing page is clearly AI generated with the “each layer a model of its own” stuff, makes little sense. The more I see AI generated copy the less it makes sense.
-
-
4
Commenter emphasizes dataset quality as critical to tool-calling performance
A developer with prior experience in similar projects stressed that dataset quality is as important as model architecture for reliable tool calling, noting the challenge of capturing all smart home devices and user phrasing variations.
“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.”
— janalsncm -
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…
-
-
3
Users test Needle 3 demo and find tool selection inconsistent
Testing showed the model handles direct commands like "turn all the lights on/off" reliably but fails on indirect requests. One tester reported "it's too cold" turned the thermostat down while "it's cold" dimmed lights, though confidence scores on these bad responses were low.
“"turn all the lights on/off" and "it's too dark in the bathroom" worked for me, but anything less direct didn't.”
— gs17 -
"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…
1 more of the top 2 · 2 posts in this stretch
-
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!
-
-
2
Early users report significant reasoning errors in demo tool calling
Community testing of Needle 3's public demo revealed consistent misrouting of natural language commands. Users documented failures including misinterpreting "I need a wee" as a music request (genre confusion), "warm the house" as a lights brightness adjustment, and "it's cold" triggering unexpected thermostat behavior.
“"I need a wee" -> tries to play music because "wee" is a genre…”
— IanCal -
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…
2 more of the top 3 · 3 posts in this stretch
-
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…
-
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…
-
-
1
Cactus releases Needle 3 with intelligence laddering architecture
Cactus Compute announced Needle 3, an 8-29 MB foundation model built on Simple Attention Network that supports tool calling and structured JSON output. The model uses intelligence laddering—each layer a sub-network with monotonically increasing capacity—allowing developers to choose from 2-20 layer configurations. The 4-layer variant, when fine-tuned on downstream tasks for one epoch, matches DeepSeek V4 Flash performance.
“Every layer of Needle 3 is a sub-network with monotonically increasing capacity. Developers can choose the right size from the 2-layer (2L) subnetwork to 20 layers (20L).”
— Cactus Compute -
first by HN Best, 8d ago · also HN Frontpage
-
What people are saying 5 voices from 2 sites · best of 20 · verbatim
- How does Needle 3's performance on tool calling compare to other small specialized models like FunctionGemma?
- Can the model's reasoning be improved enough through better training data to compete with rule-based automation for home control?
- Sep 18
-
Tool-call + structured JSON at 8–29MB is the right bet. Returning an empty list when nothing fits beats inventing a fake tool.
-
I've seen previous needle releases running on esp32s - any idea how well this new one would run on something like the newer esp32-p4?
-
Would love to see this implemented with @guggingface/kernels for shader compilation for Webgpu.
-
Really interesting project. The intelligence laddering and on-device tool calling are especially cool. Nice work getting this running across so many platforms!
- Sep 17
-
We release Needle 3: A Sliceable 8-29MB automation foundation model that can match DeepSeek V4 Flash. One set of weights, every depth from 2 to 20 layers a model of its own, 25-121M parameters at CQ2-bit, built on our Simple Attention Networks and running locally at up to 4k