Nemotron 3.5 Bolt E4B on X
NVIDIA · launched AnnouncementEvent
The launch vibe for Nemotron 3.5 Bolt E4B was mostly positive among people who ran it on their own Macs, phones, Jetson boards and GPUs, who praised its speed and tool calls for its size; the doubts were agent runs that looped, slow phones and CPUs, and a download bigger than advertised. Of the 10 posts quoted here, 6 are positive, 2 mixed and 2 negative.
Victor Bridges-Ruiz, Ph.D. @vbridgesruizPositive
“Can confirm this works on my Samsung Galaxy S20+ with the PocketPal app. Decent speed at 4.3 tokens per second.”
GooGZ AI @PaulGugMixed
“in my PA Core agent tests it didn't finish a single PA1 or PA4 run (0 of 6). … GGUF Q4_K_M with reasoning off is slower (19.2 tok/s), yet it completed all 9 PA runs and topped my 4B group!”
Spawn Lau @SpawnLauvNegative
“After running Bolt E4B on my phone, I feel we’ve still got a way to go with mobile LLMs. … The phone gets laggy and the battery drains fast.”
J @CCIE4KNegative
“Bolt, default tag settings: 1.34 tok/s … Gemma, default tag settings: 3.38 tok/s”
Dragos Roua @dragosrouaPositive
“a surprisingly capable model. I tested it on my 16GB MacBookPro and I think it’s in top 3 local models for that configuration.”
Daniel Lougen @DJLougenPositive
“Nemotron 3.5 Bolt E4B on an 8GB Jetson Orin Nano: 14.5 → 49.8 tok/s decode”
BennyDaBall @BennyDaBall_OGPositive
“Math hits 688 tok/s! The fastest setup runs in 5.8 GiB of VRAM.”
Prompt Engineer 48 @prompt48Mixed
“My read: good for a local voice assistant or a game character at 8 GB. For an agent that handles money or policy, put a calculator in the loop and verify the logic.”
Pratik Gandhi @pratikgPositive
“63 tok/s generation … 20/20 math, 10/10 code and it still pulls a hidden fact out of 32k tokens”
This page quotes only posts from people on X who said they had tested or used Nemotron 3.5 Bolt E4B.