Elon Musk's Grok 4 erhält in der späten Nacht ein großes Upgrade: 2 Millionen überragende Kontextwörter, fünfmal die "Gedächtnisleistung" von GPT-5

The Grok family has two major upgrades overnight: Grok 4 Fast extends the context to 2M, and Grok Imagine generates images so realistic that they are almost indistinguishable from real ones.

It's so fast!

Within one day, Grok has received two major updates - both Grok 4 Fast and Grok Imagine have been significantly upgraded.

Grok 4 Fast has increased the context window to 2M and raised the completion rate to 94.1% (inference) and 97.9% (non - inference).

This means that you no longer have to break up a book or an entire codebase and feed it to the model piece by piece. It can handle them all at once and stably provide results.

In resonance with this, the API call share of x.ai on OpenRouter is also on the rise, indicating that the quality of the Grok model is favored by developers.

The upgrade of Grok Imagine is mainly reflected in the generation quality - its output is so realistic that it's almost impossible to tell the difference!

2M context: Fit two copies of "War and Peace" at once

First, let's look at Grok 4 Fast. Its biggest upgrade is the ability to handle a context of 2 million tokens.

This means it can process a text volume equivalent to 1.5 million English words or 6000 pages at one time.

This scale is amazing!

It's twice that of Gemini 2.5 Pro and five times that of GPT - 5!

Without exaggeration, the 2 - million - token context window has completely changed the game rules -

An entire book, a complete codebase, or a massive dataset can all be processed at once.

You know, Leo Tolstoy's epic "War and Peace" only has about 800,000 tokens - Grok 4 Fast can easily fit two copies with room to spare!

Coupled with the lightning - fast performance of Grok 4 Fast with almost zero memory leakage, the model can conduct large - scale thinking without losing focus.

This will redefine the appearance of "real - time AI inference".

Simultaneous improvement in inference quality

While the context quantity is increasing, the inference quality of Grok 4 Fast has also significantly improved.

Specifically, in the inference mode, the completion rate of Grok 4 Fast has jumped from 77.5% to 94.1%.

In the non - inference mode, the completion rate has jumped from 77.9% to 97.9%.

This means that the probability of the model successfully outputting results that meet the requirements has increased significantly, and the failure rate (such as disconnection midway, broken output logic, or inability to complete tasks) has been significantly reduced.

It can be said that Grok has become much smarter.

In response, a netizen commented that if Grok continues to develop at this rate, "it may become the first model to truly achieve coordinated optimization of speed and inference ability rather than mutual restriction."

This is also reflected in the change of the market share of OpenRouter API calls.

According to the latest data, the API call volume of x.ai has reached 26.4%, far exceeding that of Google and Anthropic.

It can be said that it has an absolute dominant position.

Grok Imagine: So realistic that it's almost indistinguishable from real ones

Along with Grok 4 Fast, Grok Imagine has also been upgraded.

After the upgrade, the output of Grok Imagine is so realistic that it's almost impossible to tell the difference from real images.

Elon Musk himself is also having a great time playing with it on X.

Prompt: add a boyfriend and they transition into muppets

Some netizens have even played with classic plots from Western classical literature and mythology - Aeneas' escape from Troy.

This story is from the epic "The Aeneid" by the ancient Roman poet Virgil and also echoes the subsequent context of the Trojan War in Greek mythology.

There are also images from different angles -

People have used their imagination and even generated the tragic moment when Socrates drank the hemlock.

Students with ideas should go and try it out.

Reference materials:

https://x.com/BrianRoemmele/status/1987222897174380719?s=20

https://x.com/cb_doge/status/1987088612115067291?s=20 https://x.com/brianroemmele/status/1986949680278937736

https://x.com/elonmusk

https://x.com/tetsuoai/status/1987128602211864980

https://x.com/BrianRoemmele/status/1987222897174380719?s=20

This article is from the WeChat public account "New Intelligence Yuan". The author is Rhino. It is published by 36Kr with authorization.