Cactus Needle 3 Claims DeepSeek-Level Tool Calls at 29 Megabytes
Hardware / analysis
Cactus Needle 3 Claims DeepSeek-Level Tool Calls at 29 Megabytes
Hacker News commenters who tried the tiny model on ordinary requests got it right roughly a third of the time, and the company's own co-founder didn't fully disagree.
Cactus Compute says a fine-tuned subnetwork of its new Needle 3 model, shipped as a single binary of 8 to 29 megabytes, matches DeepSeek V4 Flash's performance on a tool-calling task starting at just 29 million parameters, small enough that the whole thing fits inside a phone's storage with room to spare. The company's README calls the result trading general chat ability to "beat models 10x its size on mobile tool calls and match 2-3x bigger models on extraction," a claim aimed at devices too small or too offline to call an API: phones, wearables, smart-home hubs, cars and microcontrollers.
The architecture is built around that denominator. Needle 3 uses what Cactus calls a Laddered Simple Attention Network, in which every depth from 2 to 20 layers produces its own deployable subnetwork from one shared set of weights, quantized to roughly 2.125 bits per parameter under a scheme Cactus calls Cactus Quants. That is how a single 121-million-parameter checkpoint becomes a family of binaries from 8 to 29 megabytes: a smart speaker gets the 2-layer, 8-megabyte version, and a car's infotainment unit gets the fuller 20-layer one, without retraining anything. Cactus says it trained the family on 360 billion tokens of a proprietary structured dataset, ships it Apache 2.0 licensed, and had drawn roughly 11,400 GitHub stars by Sept. 19, a day after publishing the weights on Hugging Face on Sept. 18.
On the hardware Cactus highlights most, a Raspberry Pi 5, the company's own published figures claim decode speeds from 400 to 4,000 tokens per second and prefill from 1,000 to 10,000 tokens per second, depending on which layer depth is running. Those numbers are vendor-supplied, and Cactus has not named an independent party that reproduced them.
What happened when outsiders tried it
Within hours of the Sept. 18 Show HN post, commenters ran their own tests and reported a gap between the marketing claim and the model's behavior on ordinary requests. A commenter using the handle poly2it typed "My car crashed I need help" into the demo; the model's own reasoning trace read "'crashed' implies need for music. 'play_music' with query 'car crashed'" and it queued a song, with a confidence score of 1, or 100 percent, attached to the call. Another commenter, IanCal, asked the demo to respond to being unable to see in a dark bathroom and got a vacuum switched on, a living room lit to full brightness and a track titled "wee" started instead of a light; IanCal summarized the result as leaving "the vacuum ... in a dark bathroom, the living room 100% brightness and playing 'wee'. At least there's coffee." A third commenter, gs17, said asking the model to make a room warmer instead turned a thermostat down, from 20 degrees to 18, and that asking it to check email opened a browser error or created a note containing the word "email" rather than opening a mail app.
The one number Cactus's own claim didn't survive
A commenter using the handle Scaevolus posted a head-to-head table from a tool-calling interface built against a public game database. A fine-tuned FunctionGemma model matched the correct tool shape 90.9 percent of the time and got exact arguments right 85.2 percent of the time; a fine-tuned Needle 3 running at 4-bit weights and 8-bit activations matched the tool shape only 32.2 percent of the time and got exact arguments right just 20.4 percent of the time, behind even Needle 2's prior generation on the same task. A fourth commenter, viccis, found the model set a 25-minute timer's duration field to 25 rather than 1,500 seconds, and did so while reporting 100 percent confidence in the answer.
What Cactus's own co-founder conceded
Henry Ndubuaku, Cactus Compute's co-founder and chief technology officer according to his Forbes Technology Council profile, replied directly to the criticism in the same thread: "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." That reads closer to an acknowledgment of the gap than a defense of the headline claim, and it echoes what viccis had already flagged: Needle 3 needs a narrow task and a carefully written tool definition to behave, which is a different product than the drop-in replacement the README describes.
What would change this read is an independent, published evaluation harness rather than one commenter's hand-built test against a single game database; Cactus has not pointed to one, and the DeepSeek V4 Flash comparison in its own README covers only "that task," not the tool-calling behavior outsiders tried on their own devices. Until then, the more honest comparison sits next to two other footprint-first releases this month: Edge0's 35-billion-parameter model running in under 3GB of memory and Colibri's 744-billion-parameter model streamed through 25GB of RAM. Both trade speed for a smaller footprint too, and neither has yet had to answer for what happens when a user says their car crashed.
| Correct tool shape (RuneScape task) | Model | Accuracy |
|---|---|---|
| Baseline | FunctionGemma fine-tune, BF16 | 90.9% |
| New release | Needle 3 fine-tuned, W4A8 | 32.2% |
| Prior generation | Needle 2 fine-tuned, W4 | 25.7% |
- FunctionGemma (BF16)90.9 %
- Needle 3 (W4A8)32.2 %
- Needle 2 (W4)25.7 %
Source: Hacker News comment by Scaevolus, Show HN thread, accessed 2026-09-19
Sources
More in Hardware
- 01PrismML's Ternary Bonsai 2 27B Hides a 75% ScoreThe 98.2% headline retention figure covers 20 benchmarks; on the two that test multi-step tool use, the compressed model keeps only three-quarters of its baseline score.
- 02OpenAI's Jalapeño Chip Claims Hide an Unequal TestEE Times found that OpenAI's efficiency numbers for its Broadcom-built inference chip skip an optimization rival systems are usually measured with, one worth another 3 to 5 times on its own.
- 03Icarus Robotics' JOY Robot Completes 66 of 80 ParabolasFour flights over three days in Canada gave the ISS-bound robot 22 minutes of weightlessness, using a foreign aircraft because no U.S. operator was available.
- 04Porsche's Wireless Charger Reaches Series Production at 11 kWThe Cayenne Electric is the first Porsche with the receiver built in, though the company still hasn't said what the floor plate costs or when the U.S. gets one.