LocalClaw/LLM/Needle 3
Open-weight local LLM

Needle 3

Use case: Fast chat. Check the memory fit and choose a setup path below.

1 GB catalogue minimum1 GB RAMCQ2 .cactFast chat
Choose an app

This model needs a special runtime. Unsupported apps are clearly marked.

Compare models
LM StudioNot available for this model
UnslothNot available for this model
OllamaNot available for this model
Open on Hugging FaceOfficial runtime files, licence and model card
llama.cppNot available for this model
Use with LocalClawOptional workspace after the model is installed
This model needs its official runtimeCactus Needle runtime
Parameters
121M laddered SAN
Minimum RAM
1 GB
Model size
0.029 GB
Quantization
CQ2 .cact

About Needle 3

Catalogue summary: Cactus Compute Apache 2.0 foundation model for tiny on-device tool calling, extraction and embeddings. Official `.cact` runtime artifacts run through the cactus-needle Python package, C API, browser/WASI and platform runners instead of stock GGUF or LM Studio.

Needle 3 has a catalogue minimum of 1 GB RAM with CQ2 .cact. Actual memory use and speed vary by context length, runtime, backend and system headroom.

Use the official Cactus Needle runtime setup. LocalClaw verified an official local runtime repository, not a stock GGUF or one-click LM Studio install.

Model sourcePublic official runtime repository (verified 2026-11-06): Cactus-Compute/needle3Upstream releasePaper or model cardLicence text
tool-callingagenticedgespeedembeddingstructured-output

Install path

01
Check RAM fitMinimum 1 GB RAM. Start with the CQ2 .cact quant.
02
Load the modelFollow the official Cactus Needle runtime instructions. Stock LM Studio support is not confirmed.
03
Try your workloadStart with a short prompt, check memory use and adjust context length before running longer tasks.

Catalogue record

  • Family: needle
  • Parameters: 121M laddered SAN
  • Recommended quantization: CQ2 .cact
  • Catalogue minimum RAM: 1 GB
  • Catalogue model size: 0.029 GB
  • Tags: tool-calling, agentic, edge, speed, embedding, structured-output

Practical limits

  • Catalogue RAM is a minimum estimate, not a guarantee for every context length or runtime.
  • Speed and memory use vary by quantization, backend, context length and system headroom.
  • Verify architecture, licence and usage restrictions in the linked upstream material before deployment.

Reading the scores

LocalClaw /10 is an editorial catalogue rating. Community /5 comes from member reviews. Your machine fit is a separate memory and runtime assessment.

Capability profile

Repository catalogue ratings used by LocalClaw's editorial rubric. They are not a standardized third-party benchmark.

speed
10
quality
7
coding
5
reasoning
6

Technical notes

Developer
Cactus Compute
License
Apache 2.0
Context window
1,024 tokens
Architecture
Laddered Simple Attention Network with deployable subnetworks from 2 to 20 layers, 29M to 121M parameters, CQ2-bit `.cact` runtime binaries and a byte-level grammar for exact tool-call or extraction JSON.

This model fits these next steps

Hardware fit is based on LocalClaw's RAM tier, model size and quantization metadata. Always leave memory headroom for your OS and runtime.

Entry laptop fitMacBook Air 8GBMore headroomMac mini M4 16GBAll compatible picks8GB RAM guide

Related catalogue entries

Linked mechanically by family, shared tags and nearby RAM tier; this is not a quality ranking.

Ornith 1.0 9B GGUF 9BBonsai 2 27B 27.36B (ternary)Bonsai 27B 27.3B (ternary / 1-bit)Qwen 3.5 (0.8B) 0.8BGranite 3.3 (2B Instruct) 2BGranite 4.1 (3B) 3BMiniCPM5 1B 1BQwen 3.5 (2B) 2B

Where to go next