Breaking

Gemma 4 26B Demos Local Execution of 10 Parallel Sub-Agents

June 17, 2026 at 16:14 EDT

Foundation Models
AI Agents
Open Source

On June 17, 2026, Google showcased a demo in which its open-weight model "Gemma 4 26B" orchestrated 10 parallel sub-agents locally to generate the code for an SVG art gallery within seconds, reportedly exceeding 100 tokens per second (tok/s) in generation speed.

April 2026 · Google DeepMind

Gemma 4 26B Runs Local Agentic Coding — No Cloud Required

An open-weight MoE model orchestrating 10 parallel sub-agents on-device, generating an SVG art gallery in seconds while reportedly exceeding 100 tokens per second.

25.2B

Total params (MoE)

~3.8B

Active params at inference

256K

Context window (tokens)

Parallel sub-agents in demo

Compute footprint: only a fraction stays active

MoE loads all 25.2B parameters into memory, but only ~3.8B fire per token — performance near a dense 31B class at lower compute.

25.2B

total (in memory)

~3.8B

active per token

Benchmarks (Thinking mode)

τ2-bench Retail · agentic tool use85.5%

AIME 202688.3%

MMMLU82.6%

LiveCodeBench v677.1%

Why it lands

Apache 2.0 — free commercial use & fine-tuning
Strong local agentic coding and tool calling
Runs on Ollama, llama.cpp, LM Studio, vLLM
~14.4GB at Q4_0 — privacy-focused, no cloud

The trade-offs

MoE loads full 26B-size memory despite ~4B active
Real-world ~25–40 tok/s on RTX 4060 (8GB, Q4)
Quality gaps reported in some quantized builds
Task-specific tuning sometimes needed

Continue reading

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