A 45M parameter model competing with 230M–270M models sounds ridiculous.
But that’s exactly what Needle 2 claims to do.
Needle 2:
• 45M parameters
• 14MB single binary
• 2-bit quantized
• Built for tool calling, device use, and structured extraction
Compared with FunctionGemma 270M, LFM2.5 230M, and Apple’s on-device foundation model, Needle 2 reportedly trades wins while using a footprint that can be 5–70× smaller than FP16 alternatives.
That may be the real story here.
The future of edge AI may not be about making models bigger.
It may be about making them small enough to run everywhere.
#AI #LLM #EdgeAI #OnDeviceAI #SmallLanguageModels
What is Jev?
Imagine you have a super-fast AI helper.
You give it some information and ask a simple question like:
“Is this spam?”
“Which group does this belong to?”
“Should we do A or B?”
Jev looks at the information and makes a decision.
ChatGPT is good at talking and writing.
Jev is designed to be good at choosing and judging.
It’s kind of like an AI-powered if statement for your program. ⚡
Input → Jev → Decision
Simple, fast, and easy for software to use.
#AI #Jev #Programming #ArtificialIntelligence