Intel / article

Cactus Needle 3: 8-29MB automation models can match DeepSeek V4 Flash

Source: cactuscompute.com
Author: HenryNdubuaku
Date: Published Sep 18, 2026

Context

Cactus's Needle 3 targets small-device automation by translating a text request into calls to functions an application supplies. Developers describe the available functions and their arguments; the model selects calls and fills in the required values, returning an empty list if none fits. It can also extract structured fields from messy text. One set of weights contains usable models at every depth from 2 to 20 layers, so developers can pick a size for their device, and tuning on a narrow task improves each of them. Cactus claims that after tuning on downstream tasks, models from four layers up pass DeepSeek V4 Flash. That comparison ran the Needle models against a cloud-served DeepSeek, and the page says Needle trades general chat ability for these task results, so it applies to those benchmarks.

Key takeaways · AI-distilled

Needle 3 ships as 8-29MB binaries (25-121M parameters at 2-bit) using an 'Intelligence Laddering' design where every layer from 2 to 20 is itself a deployable subnetwork from one set of weights, rather than training separate model sizes.
Its MLP replaces the standard dense feedforward layer with a 'Monarch Hadamard MLP' — Walsh-Hadamard-initialized Kronecker factor pairs plus per-channel scales and a rank-8 gate — giving roughly O(d√d) compute per versus the O(d²) a dense 4x-expansion MLP would cost.
On a Mobile Actions benchmark (phone commands scored on exact tool call), the 20-layer Needle 3 scores 86.0 versus LFM2.5 1.2B at 82.4, Qwen3.5 0.8B at 76.0, and Apple's on-device model at 57.6 — rivals evaluated at full f16 against Needle's shipped 2-bit binary.
Every response carries a confidence score (the minimum of a judgment on the finished call and its decode probability), letting a deployment act automatically above a chosen threshold, ask for confirmation below it, or escalate to a bigger model.
The team reports that just a 4-layer Needle on a narrow task can reach DeepSeek V4 Flash-level performance on that task, though they stress this only holds for narrow, not general-purpose, tasks.

Terms in this piece · Glossary

quantization — Shrinking a model by storing its numbers less precisely — like rounding — so it runs faster and fits on smaller hardware, at a small quality cost.
token — The chunk of text a model reads and writes in — roughly three-quarters of a word — and the unit AI usage is billed in.
calibration — How well a model's confidence matches reality — a calibrated model saying "90% sure" is right about 90% of the time.
fine-tuning — Taking a trained model and training it a bit more on your own examples so it gets better at one specific job.

Read the source cactuscompute.com