Google has started its 'most ambitious pre-training run yet' for Gemini 4. The announcement came via a post on X from @OfficialLoganK, retweeted by @mweinbach, with no further details disclosed.
Key facts
- Gemini 4 pretraining started per @OfficialLoganK post.
- No compute, dataset, or timeline disclosed.
- Prior release: Gemini 1.5 Pro in early 2025.
- TPU v6 could enable 10^26 FLOPs training runs.
- GPT-5 and Claude 4 are competing frontier models.
Google has started its 'most ambitious pre-training run yet' for Gemini 4, according to a post on X from @OfficialLoganK, retweeted by @mweinbach According to @mweinbach. The post offers no details on compute scale, dataset size, or expected release date, leaving the scope of 'most ambitious' undefined against the backdrop of Google's prior models.
The last public milestone was the Gemini 1.5 Pro release in early 2025, which introduced a 1-million-token context window. Google has since faced pressure from OpenAI's GPT-4o and Anthropic's Claude 3.5 Opus on multimodal and reasoning benchmarks. Starting pretraining now suggests a target release in late 2026 or early 2027, consistent with Google's rough annual cadence for major Gemini versions (Gemini 1.0 in Dec 2023, Gemini 2.0 in Dec 2024).
The term 'most ambitious' likely implies larger training compute and dataset size, but without specific numbers—flops, parameters, tokens—the claim remains a marketing signal. Competitors like OpenAI and Anthropic have disclosed training costs and scale for their frontier models; Google has not yet matched that transparency.
What 'Most Ambitious' Might Mean
If Google scales its TPU v6 deployments—announced in 2025 with up to 4x performance per watt over v5—Gemini 4 could train on 10^26 FLOPs or more, on par with GPT-4-class runs. The post's vagueness leaves room for speculation on architectural innovations, such as deeper mixture-of-experts layers or novel attention mechanisms, but no evidence supports those conjectures yet.
Competitive Context
OpenAI's GPT-5 is rumored to be in late-stage training, while Anthropic's Claude 4 is expected in mid-2026. Google's early start for Gemini 4 could signal an attempt to regain leadership in multimodal reasoning, where GPT-4o currently holds a narrow edge on MMLU-Pro and MATH benchmarks. The lack of a benchmark target in the announcement is notable—Google often cites specific metrics for major releases.
Industry Skepticism
The announcement's brevity has drawn skepticism from AI analysts, who note that 'most ambitious' is a subjective claim without disclosed compute or data. Google has a history of overpromising on Gemini—the demo video for Gemini 1.0 was criticized for being aspirational rather than real-time. Without concrete details, this post is a signal of intent, not a milestone.
Key Takeaways
- Google starts Gemini 4 pretraining, its most ambitious run yet.
- No details on compute or timeline; competitive pressure from OpenAI and Anthropic.
What to watch
Watch for Google to disclose training compute and dataset size in a future blog post, and for benchmark scores on MMLU-Pro, MATH, and multimodal tasks in late 2026. Also monitor TPU v6 deployment updates from Google Cloud.
