Needle 3: an 8–29 MB model with flexible depth

What Cactus Compute actually released

I would describe Needle 3 not as a long-context champion, but as an extremely compact model whose depth can scale for local automation. In the project repository and its Hugging Face model card, Cactus Compute calls it a Laddered Simple Attention Network and lists deployable subnets from 2 to 20 layers. Its stated footprint ranges from 8 to 29 MB.

Its internals include Monarch Hadamard MLPs, GQA attention with causal convolution branches, n-gram engram memory read through gather operations, and multichannel hyperconnections. This does not look like a random collection of tricks: the apparent goal is to keep the model small while making it usable across different compute budgets. The ability to choose depth matters more here than a single fixed model size.

The official positioning is also narrow and practical: tool calling, structured extraction, mobile actions, and local schema-constrained agents. Package documentation covers the Needle class, response format, tool use, tuned weights, a command-line interface, and offline mode. In other words, this release should primarily be assessed as a component for devices and automation, not as a general-purpose large model.

As of September 19, 2026, the available official materials do not show a long-context benchmark table or needle-in-a-haystack results specifically for Needle 3. That is an inconvenient but important gap: the name may invite assumptions about strong retrieval from a large context, yet the published material does not support that conclusion. The architecture is interesting; extraction quality at substantial context depths remains an open question.

What changes for local agents

The key change is that developers receive not one rigid model variant, but a family of deployable subnets within a shared architecture. This is useful when the same agent must fit different device constraints. Scenarios involving local tool calls and predictable structured output stand to benefit most.

I would test three things before admiring the architecture diagram: tool-selection accuracy, compliance with the requested schema, and degradation as the number of layers falls. The n-gram memory also needs a separate test: it is important to learn whether it helps on real recurring patterns or merely adds another mechanism with its own edge cases.

For that reason, the release looks more substantial than a typical parameter-count race, but it does not prove a user-facing link to efficient long-context processing. Needle 3 already has a clear engineering shape, while a convincing public picture of its limits is still missing.

We previously reviewed Pony Alpha on OpenRouter, a model with a 200K-token context window for testing architectures and workflows. This example helps compare Needle 3's approach to long-context tasks with models already available on the market.