Breaking

Musk Says xAI's 2-Trillion-Parameter Grok Model Will Finish Initial Training Next Week

July 18, 2026 at 07:38 EDT

Foundation Models
Infra & Chips

Elon Musk says xAI's next flagship, a 2-trillion-parameter Grok model, will complete initial training next week and could surpass Moonshot AI's Kimi while retaining the speed and token efficiency of the current 1.5-trillion-parameter Grok 4.5. The disclosure signals that xAI is pushing ahead with an aggressive scaling cadence just weeks after shipping its latest flagship.

xAI Roadmap · Frontier Model Race

A 2-Trillion-Parameter Grok Finishes Initial Training Next Week

Elon Musk says xAI's next flagship is "better than our 1.5T in every way" — and could surpass Moonshot's Kimi K3 while keeping the speed and token efficiency of Grok 4.5.

Parameters in the upcoming Grok flagship

~80TPS

Grok 4.5 throughput the new model aims to preserve

4.2×

Fewer output tokens vs Opus 4.8 on SWE-Bench Pro

Scale, compared

Parameter count — the contenders

1.5T

Grok 4.5
current flagship

Grok 2T
upcoming

~2.8T

Kimi K3
open MoE

The pitch: match or beat the larger 2.8T Kimi K3 while running at the speed and cost profile of the smaller 1.5T Grok 4.5.

Grok 4.5 — the efficiency baseline

500k-token context window
~$2 / 1M input · ~$6 / 1M output tokens
~80 tokens/sec throughput
Strong coding & agent-task scores

Kimi K3 — the target to beat

~2.8T open mixture-of-experts
1M-token context, native multimodal
Tops some front-end coding boards
79% success rate in one repair harness

Still unconfirmed

Only initial training finishing next week is stated. Public release date, pricing, context length, and full benchmark results for the 2T model have not been disclosed — any September launch window is speculative.

Continue reading

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