NVIDIA has demonstrated a robot running Gemma 4 E4B, a model from Google's Gemma family, on a single 8GB Jetson module, combining live conversation, tool calls, animations and a vision model on one compact board.
NVIDIA · Gemma 4 E4B · Embodied AI
A robot runs a multimodal Gemma model on one 8GB board
NVIDIA demonstrated live conversation, tool calls, animation control and a vision model running together on a single compact Jetson-class module, letting a robot scan a room, identify a large television and start a movie break.
8 GB
Memory on a single module
4
Functions running together: conversation, tool calls, animation, vision
1
Compact board doing the work, without a constant cloud link
Memory budget: the whole 8GB module
Each block = 1 GB of the module's total memory. The split per function was not disclosed.
8 GB
Four AI functions and a vision model share one small memory pool, which is the practical test for moving AI into robots, appliances and industrial equipment.
The demo loop: perceive, decide, act
→
STEP 2
Identify a large television
→
STEP 3
Start a movie break via tool call
What stands out
Vision, language and actions run on one small board
Tool calls let the model act on its environment, not just generate text
Less dependence on a constant cloud connection
What remains unproven
Showcase only, with no independent evaluation yet
Reliability, response speed and safety in real conditions are unknown
Whether developers can reproduce it at scale
Not yet disclosed
Performance, power, price and availability remain open questions.
Latency
Benchmarks
Power consumption
Pricing
Continue reading The rest of this article is for AI News Blitz readers. Choose an option below to keep reading.
Already purchased? Sign in ✓ Signed in — this article isn’t included in your current plan.Unlocking the full article...