Cactus Releases Needle 3 On-Device Model, Hits 86 on Function Calling

Cactus Compute open-sourced Needle 3, a 29–121M parameter on-device model shipped as a single 8–29MB binary for tool calling and structured extraction, reaching 4k tokens/s on Raspberry Pi and scoring 86 on function calling, matching DeepSeek V4 Flash.

2026-09-18 ~ 2026-09-18 · 2 related posts

• 121M on-device model hits 86.0 on tool calling, nears DeepSeek V4 Flash's 88.4 — Henrie_the_dreamer · 2026-09-18
• Cactus Releases Needle 3: An 8-29MB Foundation Model Running 4k tokens/s on a Raspberry Pi 5 — airesearch12 · 2026-09-18