Breaking

Google Launches Agent-Collaboration Challenge to Speed Up Gemma 4 E4B on Hugging Face

June 9, 2026 at 11:53 EDT

Foundation Models
Open Source
AI Agents

On June 9, 2026, Google announced the "Fast Gemma Challenge," an effort in which multiple AI agents collaborate on Hugging Face to make inference of the lightweight open-weight model "Gemma 4 E4B" faster. Over several days, dozens of agents will take part, and participants can submit their own agents. A dedicated dashboard to track progress has been made available on Hugging Face.

June 9, 2026 · Google & Hugging Face

The Fast Gemma Challenge: AI Agents Race to Speed Up Gemma 4 E4B

Dozens of AI agents will collaborate over several days, sharing resources to push local inference on the lightweight E4B model even faster. Autoresearch, the team says, is far more fun when it is collaborative — though no concrete results are out yet.

~4.5B

Effective parameters (~8B with embeddings)

128k

Context length, multimodal & audio-native

~6GB

RAM enough for search, code & cited answers

Inference memory drops sharply with quantization

Approximate memory needed to run E4B locally

17.9GB

FP16

8.9GB

INT8

INT8 roughly halves the footprint — already fast, the challenge aims to make it faster still.

E4B benchmark scores

MMMLU

69.4%

MMMU Pro

52.6%

LiveCodeBench

52.0%

AIME 2026

42.5%

Praise

"Surprisingly fast and intelligent" — called by some the best model for local AI, running everyday tasks and sub-agent roles on modest RAM.

Caveats

Falls short of larger models like GPT-5.4 on tasks such as summarization, with gaps in complex reasoning and multilingual work — best as a supplement.

How the challenge works

Dozens of agents join

→

Share resources

→

Optimize local inference

→

Faster E4B

Built on Hugging Face's ecosystem — transformers, llama.cpp, MLX, WebGPU — with progress expected to be updated over the coming days.

Continue reading

The rest of this article is for AI News Blitz readers. Choose an option below to keep reading.