ai-wars
xAI's Timeline Dates Grok to November 3, 2023. The Open Weights Are a 314 Billion Parameter Model.
The announcing post describes Grok-1 and a 33 billion parameter Grok-0. The open-release post says Apache 2.0, 314 billion parameters, and a pretraining run that concluded in October 2023. The company timeline dates that release to March 17, 2024. Later model posts carry their own dates where the article prints one.
An announcement, an open-weight checkpoint, and a later model with a new number are different releases. A benchmark in a company post is the company’s score. A date on the company timeline is the timeline’s date when the article body does not print one.
The product hub is xAI and Grok. The cluster named in the later posts is Colossus. Where the companies sit in the filings is who owns X, xAI, and SpaceX.
What’s confirmed
November 3, 2023, on the company timeline: Announcing Grok
x.ai/company, fetched October 20, 2026, lists under 2023: “Announcing Grok Nov 3.” The article, Announcing Grok, does not print a calendar date in the body this desk read. The page header fetched that day says SpaceXSI. The sign-off in the body is “Thank you, the xAI Team.”
The post says Grok is “an AI modeled after the Hitchhiker’s Guide to the Galaxy” and “is still a very early beta product – the best we could do with 2 months of training.” It says a fundamental advantage is “real-time knowledge of the world via the 𝕏 platform.” It says the engine is Grok-1, “our frontier LLM, which we developed over the last four months.” It says that after announcing xAI the team trained a prototype, Grok-0, with 33 billion parameters. It says Grok-1 achieved 63.2% on HumanEval and 73% on MMLU in the prose, and the table prints Grok-1 at 62.9% on GSM8k (8-shot), 73.0% on MMLU (5-shot), 63.2% on HumanEval (0-shot), and 23.9% on MATH (4-shot). It says Grok passed the 2023 Hungarian national high school finals in mathematics with a C (59%), against Claude 2 at 55% and GPT-4 at 68%, all at temperature 0.1.
Label: The company’s post, dated by the company timeline to November 3, 2023. The percentages are the post’s. “State-of-the-art” and the comparisons with GPT-3.5, Inflection-1, and GPT-4 are the post’s. The prose says 73% on MMLU. The table says 73.0% (5-shot). This page uses the table’s cell when it cites the table.
March 17, 2024, on the timeline: the open weights
The company timeline lists “Open Release of Grok-1 Mar 17” under 2024. The open-release post does not print that date in the body. It says xAI is releasing the base model weights and network architecture of Grok-1, “a 314 billion parameter Mixture-of-Experts model trained from scratch by xAI.” It says this is the raw checkpoint from pretraining, “which concluded in October 2023,” and that the model “is not fine-tuned for any specific application, such as dialogue.” It says the weights and architecture are released under the Apache 2.0 license. It says 25% of the weights are active on a given token. It points to github.com/xai-org/grok.
The GitHub README for xai-org/grok-1, fetched October 20, 2026, says the repository contains JAX example code for the Grok-1 open-weights model. It says the model is 314B parameters, a mixture of 8 experts, with 2 experts used per token, 64 layers, 48 attention heads for queries and 8 for keys and values, embedding size 6,144, a SentencePiece tokenizer with 131,072 tokens, and a maximum sequence length of 8,192 tokens. It says: “The code and associated Grok-1 weights in this release are licensed under the Apache 2.0 license. The license only applies to the source files in this repository and the model weights of Grok-1.”
A direct fetch of https://raw.githubusercontent.com/xai-org/grok-1/main/LICENSE returned 404 the same day. The license sentence this page uses is the README’s, and the Apache 2.0 sentence in the open-release post.
Label: xAI’s post and the README. 314 billion, October 2023, not fine-tuned, Apache 2.0, and 25 percent active are the post. Eight experts and two per token are the README. March 17, 2024, is the company timeline. The article body does not print it.
Later model posts, as the timeline and the articles date them
The same company page dates further model lines. Where this desk also opened the article, the article’s own claims stay with the article.
| What the company timeline prints | Date on that timeline | What this desk opened |
|---|---|---|
| Announcing Grok-1.5 | Mar 28 | The post. No date in the body. It says Grok-1.5 has a context length of 128,000 tokens and will be available to early testers and existing Grok users on X “in the coming days.” It says that releasing Grok-1’s weights “two weeks ago” showed progress through last November. It prints 50.6% on MATH, 90% on GSM8K, and 74.1% on HumanEval |
| Grok-1.5 Vision Preview | Apr 12 | The timeline line. The article was not opened |
| Grok-2 Beta Release | Aug 13 | The post. No date in the body. It announces Grok-2 and Grok-2 mini, says an early version was tested on the LMSYS leaderboard as “sus-column-r,” and says that at the time of the post it was outperforming Claude 3.5 Sonnet and GPT-4-Turbo. It says both are in beta on X |
| Grok 3 Beta | Feb 19 | The post prints “Back to news Feb 19, 2025.” It says Grok 3 was trained on Colossus “with 10x the compute of previous state-of-the-art models,” prints an Elo of 1402 in the Chatbot Arena, and says that with cons@64, Grok 3 (Think) achieved 93.3% on the 2025 AIME, released February 12. It also prints 84.6% on GPQA and 79.4% on LiveCodeBench |
| Grok 4 | Jul 9 | The post prints “Back to news Jul 9, 2025.” The headline claim is “Grok 4 is the most intelligent model in the world.” It says training used Colossus, “our 200,000 GPU cluster,” and that stack work increased training compute efficiency by 6x |
| Grok 4.1 | Nov 17 | The timeline line. The news index, fetched the same day, says Grok 4.1 is available on grok.com, X, and the iOS and Android apps, in Auto mode and as a picker choice. The article body was not opened |
| Introducing Grok 4.5 | Jul 16 | The timeline line. The index describes Grok 4.5 as “SpaceXSI’s smartest model built for coding, agentic tasks, and knowledge work.” The article body was not opened |
| Introducing Grok 4.6 | Aug 12 | The timeline line. The index says Grok 4.6 “builds on Grok 4.5 with a particular focus on long-running agents.” The article body was not opened |
| Introducing Grok 4.7 | Sep 21 | The timeline line, “Our most capable reasoning model,” on the company page. The article body was not opened |
The news index also lists many product posts that are not a new flagship model: image and voice releases, API availability, and Grok Bot. Those titles are on the index. They are not given their own rows here.
Label: Company claims, each tied to the page that prints them. “Most intelligent” and “smartest” are the posts’ and the index’s words. The 128,000-token figure is the Grok-1.5 post. The 8,192-token figure is the Grok-1 README. They describe different releases. “Two weeks ago” is the Grok-1.5 post’s phrase for the weight release. The timeline dates that release to March 17 and Grok-1.5 to March 28. This page does not convert “two weeks” into a day count.
The Grok hub already records a July 8, 2026 homepage mention of Grok 4.5. The company timeline fetched on October 20, 2026, dates “Introducing Grok 4.5” to July 16. Those are two dates already on this site’s sources. This page does not average them.
Claims and open questions
The November 3 and March 17 dates are the company timeline. The announcement and the open-release article, as fetched, do not print those dates in the body.
The LICENSE file URL returned 404. The README and the open-release post both say Apache 2.0. This page does not quote a license text it did not retrieve.
Benchmark percentages are the company’s reported scores. They are not an independent evaluation this desk ran. The Grok-2 post’s leaderboard sentence is the post’s claim about a point in time.
Grok 4.1, 4.5, 4.6, and 4.7 are dated here from the company timeline and the news index. Their articles were not opened, so this page does not repeat parameter counts or benchmark tables for them.
On the map
xAI and Grok is the product hub. Colossus is the cluster the Grok 3 and Grok 4 posts name. Who owns X, xAI, and SpaceX is the filing record. The AI Wars map is the section.
Sources
Checked October 20, 2026.
| Source | Grade | What it supports |
|---|---|---|
| x.ai/company | A, company timeline | Nov 3, 2023. Mar 17, Mar 28, Apr 12, and Aug 13, 2024. Feb 19 and Jul 9 and Nov 17, 2025. Jul 16, Aug 12, and Sep 21, 2026. As the timeline prints those lines |
| Announcing Grok | A, company post | Grok-0 at 33 billion parameters. Grok-1 table: 62.9% GSM8k, 73.0% MMLU, 63.2% HumanEval, 23.9% MATH. Hungarian exam 59%. Early beta, two months of training. Body has no calendar date |
| Open Release of Grok-1 | A, company post | 314 billion parameters. Mixture of experts. 25% of weights active. Pretraining concluded October 2023. Not fine-tuned for dialogue. Apache 2.0. Body has no calendar date |
| xai-org/grok-1 README | A, company repository | 314B. 8 experts, 2 per token. 64 layers. 48 query heads, 8 key/value heads. Embedding 6,144. 131,072-token tokenizer. Context 8,192. Apache 2.0, as the README states it. The raw LICENSE URL returned 404 |
| Announcing Grok-1.5 | A | 128,000 tokens. MATH 50.6%, GSM8K 90%, HumanEval 74.1%. “Two weeks ago” |
| Grok-2 Beta Release | A | Grok-2 and Grok-2 mini. “sus-column-r.” The post’s comparison with Claude 3.5 Sonnet and GPT-4-Turbo |
| Grok 3 Beta | A | Article date Feb 19, 2025. Colossus, “10x the compute.” Elo 1402. AIME 93.3% at cons@64. GPQA 84.6%. LiveCodeBench 79.4% |
| Grok 4 | A | Article date Jul 9, 2025. “Most intelligent model in the world,” as the post says. Colossus, 200,000 GPUs. 6x efficiency, as the post says |
| x.ai/news index | A, company index | Titles and index blurbs for Grok 4.1, Grok 4.5, and Grok 4.6. Those article bodies were not opened |