Breaking

Google DeepMind Confirms Pre-Training of Gemini 4 Has Begun, Its 'Most Ambitious' Run Yet

July 21, 2026 at 12:17 EDT

Foundation Models

Google has started pre-training its next flagship AI model, Gemini 4, a Google DeepMind staff member confirmed publicly, describing the effort as the company's most ambitious training run to date. The disclosure, made on July 21, 2026, marks the first official acknowledgment that work on the successor to the current Gemini 3.x family is underway.

July 21, 2026 · Google DeepMind

Gemini 4 Enters Pre-Training — Google's Most Ambitious Run Yet

A Google DeepMind staff lead publicly confirmed the successor to the Gemini 3.x family has begun its foundational, compute-heavy phase — a signal of direction, not yet a shippable product.

Phase 1

Pre-training only — the earliest, most compute-intensive stage

New Flash models shipped the same day (3.6 Flash, 3.5 Flash-Lite, 3.5 Flash Cyber)

~6 mo

Typical training cycle — some speculate a late-2026 arrival (unconfirmed)

Gemini 3.6 Flash: leaner output, same cost

Roughly 17% fewer output tokens than 3.5 Flash — higher quality on coding & knowledge work at the same price.

100%

3.5 Flash
output tokens

~83%

3.6 Flash
output tokens

Where Gemini 4 sits in the build pipeline

Pre-training

Broad knowledge from massive datasets — happening now

→

Post-training

Fine-tuning for tasks — not yet reached

→

Release

Usable model still months away

Optimism

Developers praised the Flash updates' token efficiency and better economics for agentic, multi-step workloads. Many voiced anticipation for the scale-up.

Skepticism

Others cited frustration with the 3.x rollout cadence and release ordering, noting a pre-training run has no immediate impact.

What Google has not disclosed:

Parameter count Training data volume Compute budget Benchmarks Pricing Release schedule

Continue reading

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