Editorial Team
6 min read
Elon Musk says xAI's 2-trillion-parameter Grok 4.6 will finish initial training next week. The claim sets a checkpoint, not a release date, and leaves too little technical or commercial detail for a sound comparison with Kimi K3.
Elon Musk has set a near-term training checkpoint for xAI's next Grok model and a competitive aspiration against Kimi K3. Neither claim yet establishes what Grok 4.6 can do, when customers can use it or whether its economics improve on the model it is meant to follow.
Musk said on July 18 that xAI's 2-trillion-parameter model was better than its 1.5-trillion-parameter model “in every way” and would finish initial training the following week. He said it might exceed Kimi while retaining speed and token efficiency close to the 1.5-trillion-parameter model he called Grok 4.5. In a separate reply, he confirmed the new model's name as Grok 4.6.
The move from 1.5 trillion to 2 trillion parameters is an increase of one-third. That arithmetic is clear; its performance significance is not. The capability, speed and efficiency statements come from Musk, and “might” makes the Kimi comparison a forecast rather than a reported result.
Two accounts published after Musk's posts said xAI had not disclosed a release date or complete technical specifications. One said training details and full metrics were absent. The other said access terms and independent results were also unavailable.
The distinction between initial training and release is material. The first account said Musk had described the 1.5-trillion-parameter V9-Medium base model in May as having finished training before fine-tuning and reinforcement learning. That sequence offers no basis for converting next week's proposed checkpoint into a Grok 4.6 launch window.
The public naming sequence is unresolved too. In January, xAI said Grok 5 was already in training in a financing announcement. The retained sources do not explain how that work relates to the model Musk now calls Grok 4.6.
Company-reported total parameter counts for Grok 4.5, Grok 4.6 and Kimi K3. Total size alone does not show active parameters or performance. Source: Elon Musk on X.
Moonshot AI describes Kimi K3 as a 2.8-trillion-parameter model with native vision and a 1-million-token context window in its technical post. Its published total is 40% larger than Grok 4.6's stated 2 trillion. Put the other way around, Grok 4.6's total would be about 29% lower; calling it 40% lower would use the wrong denominator.
Even the corrected calculation has limited value. K3 is a sparse mixture-of-experts model that, according to Moonshot, activates 16 of 896 experts per token. xAI has not disclosed whether Grok 4.6 is dense or sparse, how many parameters it activates, or what hardware and numerical precision it uses for inference. Total parameters therefore cannot establish which model does more computation per token.
Kimi's disclosure also narrows what “exceed Kimi” would mean. Moonshot says K3 still trails Claude Fable 5 and GPT 5.6 Sol overall in its own evaluation suite. Its benchmark notes use different agent harnesses for some models and include results taken from outside leaderboards as well as Kimi-run evaluations. Beating K3 on an xAI-selected test would not by itself demonstrate market leadership.
There is a distribution difference as well. K3 is already available through Kimi products and an API, and Moonshot said full model weights would be released by July 27. It also said more architecture, training and evaluation details would arrive with a later technical report. No comparable access or publication plan has been disclosed for Grok 4.6.
K3's published economics show the information a serious comparison requires. Moonshot lists API prices of $0.30 per million cache-hit input tokens, $3 per million cache-miss input tokens and $15 per million output tokens. It recommends deployments with at least 64 accelerators and warns that generation quality can become unstable when a harness fails to preserve the model's thinking history. Those are vendor disclosures, not guarantees of cost or reliability on a customer's workload, but they expose deployment variables missing from the Grok 4.6 claim.
A 2022 compute-optimal scaling study supplies a useful counterexample to simple size rankings. The researchers trained more than 400 language models, spanning 70 million to more than 16 billion parameters and 5 billion to 500 billion training tokens. Under a fixed compute budget, they found that model size and training-token count should rise together.
Their 70-billion-parameter Chinchilla model used the same compute budget as the 280-billion-parameter Gopher, trained on four times as much data and outperformed Gopher across a broad range of downstream evaluations. The study does not predict the result of a 2026 contest between Grok and a sparse Kimi model. It shows why xAI's undisclosed training-token count, data, architecture and post-training work are necessary to interpret the 2-trillion figure.
xAI has disclosed substantial capital and computing infrastructure. The company said its Series E raised $20 billion, above a $15 billion target, with Nvidia and Cisco Investments among the strategic participants. It said Colossus I and II ended 2025 with more than 1 million H100 GPU equivalents and that the financing would support infrastructure, product development and research.
Those are company disclosures, not a cost breakdown for Grok 4.6. The company described capacity across two systems in H100 GPU equivalents; it did not present that figure as a count of physical chips.
A separate study covering 500 AI supercomputers from 2019 through 2025 identified the 200,000-chip Colossus as the leading system in March 2025. Its dataset assigned the system $7 billion in hardware cost and 300 megawatts of power demand. Those March 2025 figures cover a differently defined system and cannot be subtracted from, or directly compared with, xAI's year-end H100-equivalent claim.
The figures establish that xAI can finance and operate unusually large training infrastructure. They do not answer the commercial question Musk has raised. Training capacity says little about how many accelerators Grok 4.6 will need in production, how fast it will serve requests, how many tokens it will consume to solve a task or what xAI will charge.
Finishing initial training would resolve only the checkpoint Musk announced. Four later disclosures would determine whether Grok 4.6 is a product advance rather than another larger run:
Until those arrive, the supported conclusion is narrow. Musk has identified a 2-trillion-parameter model and forecast the end of its initial training. The retained evidence does not yet show that Grok 4.6 surpasses Kimi K3, preserves Grok 4.5's efficiency or improves the economics of deploying xAI's models.
Anthropic has put Bun’s Rust port into Claude Code ahead of Bun 1.4’s general release, creating a real but tightly controlled proving ground for an AI-led migration whose total cost and broader reliability remain unsettled.
AIdeaLab has released two anime-focused Wan 2.2 video-model fine-tunes under Apache 2.0. The weights are commercially usable and locally deployable, but current benchmarks, production economics and training-data detail remain undisclosed.
Zhipu reportedly reached $1 billion in annual recurring revenue in July, roughly four times a March estimate, but the unconfirmed run rate is not annual sales and still sits far ahead of recognized cloud revenue while margins remain thin.
OpenAI strategy executive Dean Ball says open-weight AI will deter investment and end in state provision. Kimi K3 exposes a nearer contest over API prices, scarce compute, hosting revenue and regulatory barriers.
A competitor identified a concrete case in which MEDLEY-BENCH may aggregate confidence labels attached to different propositions. The retained record does not show whether the problem is widespread or changes model rankings, while Kaggle's defense of its human review does not resolve the technical question.
OpenAI's Codex 0.144.6 sets 272,000-token context metadata for GPT-5.6 Sol, Terra, and Luna after one Sol user documented a brief 372,000-token setting, renewing a dispute over how Codex accounts for input, output, and compaction headroom.
Dave Eggers told OpenAI that ChatGPT could deprive students of their writing voices. A national teen survey and a small essay-writing preprint identify real risks without proving generational harm, while schools and unions are already adopting a broader mix of AI products and training.
Anthropic workers are giving to federal candidates in unusual numbers, but the largest AI-politics balance sheets remain concentrated among executives and super PACs. The comparisons show a new donor network—not a unified bloc, proof of company direction or evidence that spending decides elections.
Reddit says upgraded automated systems block 23 million spam views daily and reduce exposure, but its disclosure does not identify how much caught spam is AI-generated or provide the denominators and error rates needed to judge the filter.
A review classified 88 of 120 repositories linked to AI-flagged Flathub submissions as abandoned, a stark result that helps explain reviewer frustration but does not measure AI’s effect on software maintenance or the ban’s impact.
IBM says customers diverted late-quarter budgets to scarce AI hardware. Its own product results show why investors now need proof that delayed mainframe and software deals were not lost priorities.
Cmpunlocker says patched Nvidia open kernel modules can expose 64GB of HBM on an 8GB CMP 170HX or 40GB on a 10GB card while removing an SM throttle. The exploit mechanism is documented, but the retained evidence does not independently validate the tool across full-memory workloads, error rates, power use or A100-class performance.
An unidentified industry source says at least one Nvidia board partner has physical RTX 50 Super cards but cannot sell them, with 3GB GDDR7 quoted at $60 to $70 per chip. The reported memory bill explains a plausible pricing problem, but neither Nvidia nor a second independent source has confirmed the products, the hold or the component terms.
Christian Faith Madison’s estate alleges GPT-4o sustained a months-long narrative of prophecy, sacrifice and resurrection before her death. Public safety disclosures and controlled research make that failure mode plausible, but neither establishes which model behavior she encountered or whether it caused her death.
Shanghai AI Laboratory has released Intern-S2-Preview-397B as downloadable weights and a hosted API, but its own 35B alternative offers the same context window and a far smaller disclosed scale while the larger model still lacks auditable comparative scores, pricing and physical-validation results.
Mixfont's free Decoy Font gives each glyph a sharp decoy and a blurred intended letter, creating a hurdle for pixel-based readers while leaving selectable text, known-image techniques and accessibility costs outside its protection.
Anthropic’s 20-year lease gives TeraWulf a customer for 401 megawatts of future AI infrastructure, but rent starts only after delivery and the proposed utility deal passes power and grid costs to TeraWulf, leaving financing and project margin unresolved.
OpenAI has put conversations and Projects back in its redesigned ChatGPT desktop app and enabled cloud Work threads to move across devices, correcting the launch's biggest usability failures without merging local Work or Codex histories.
Twenty-nine countries signed an agreement creating WAICO as an independent intergovernmental organization, while China paired the launch with capacity-building offers that are not yet confirmed as WAICO programs.
Meta reportedly plans to put departing AWS compute executive Dave Brown to work on its data-center buildout, adding hyperscale operating experience while leaving any customer-facing cloud business conditional and undefined.