A new open-weight AI from China is surpassing the leading models from OpenAI and Claude Fable.
Moonshot's Kimi K3, a colossal 2.8-trillion-parameter model, surpasses Fable 5 and GPT 5.6 Sol in certain benchmarks.
China’s Moonshot AI has introduced Kimi K3, an extensive 2.8-trillion-parameter model designed for coding, research, reasoning, and visual tasks. Moonshot acknowledges that K3 is still behind Claude Fable 5 and GPT 5.6 Sol in terms of overall performance. However, its benchmark results place it surprisingly close to both models, with K3 outperforming them in several tests.
How does Kimi K3 compare to the top closed models?
K3 achieved a score of 77.8 on Program Bench, just edging out Fable 5's score of 76.8 and GPT 5.6 Sol's 77.6. Moreover, it excelled in BrowseComp with a score of 91.2 and SWE Marathon with 42.0, surpassing both competitors.
Other results indicate that there is still progress to be made. K3 scored 67.5 on DeepSWE, while Fable 5 scored 70.0 and GPT 5.6 Sol scored 73.0. Moonshot also states that the overall user experience is still lacking compared to both proprietary models.
These comparisons come with a significant caveat. Moonshot claims that all Fable 5 results might incorporate fallback features to another model, while GPT 5.6 Sol results might be influenced by cyberguards that limit specific responses. Thus, while the comparison is valuable, it’s not entirely balanced.
Open weights put pressure on OpenAI and Anthropic
A key aspect of K3 may be its open-weight release. Moonshot intends to release the full model weights by July 27, allowing users to download and run K3 locally if they possess the necessary hardware for such a large model. Developers will also have the option to modify and fine-tune it for particular tasks.
K3 is more affordable than the premium U.S. models it competes with, yet it doesn't fall into the category of being very inexpensive by Chinese AI standards. Its API pricing stands at $0.30 per million cached input tokens, $3 for uncached input, and $15 for output, placing it closer to Anthropic's mid-tier models rather than the heavily discounted prices seen with earlier Chinese releases.
Typically, Chinese AI models are significantly cheaper than their U.S. counterparts, prompting some American startups to adopt them to reduce expenses. While Kimi K3 is not the least expensive among Chinese models, its near-top-tier performance at a lower price still compels OpenAI and Anthropic to validate their higher pricing.
I have approximately four years of experience, primarily focusing on gaming, PC hardware, and smartphones. In my spare time, I enjoy...
OpenAI updates ChatGPT desktop following user criticism over its recent redesign
ChatGPT's desktop application is receiving a vital update. Following the merger of Chat, Work, and Codex into a single integrated desktop app about a week ago, the user experience was plagued with issues, hindering basic features like chat history and complicating the mode-switching process. In response, OpenAI has implemented a series of fixes based on user feedback to ensure a seamless experience across all devices.
I have five mechanical keyboards I would enthusiastically endorse for the Back-to-School season
Selecting the right mechanical keyboard enhances every assignment, gaming session, and study time. Choosing a mechanical keyboard differs from picking a laptop; it's not just about flashy specifications or benchmark results, but rather about everyday enjoyment. Whether taking class notes, writing essays, coding late into the night, or relaxing with some games post-lecture, I've reviewed numerous options and consistently returned to these five keyboards that cater to a range from budget-friendly beginner boards to high-end enthusiast models and gaming-focused options. Most importantly, these are keyboards I would genuinely recommend for the Back-to-School period.
HP penalized millions for collusion over ink and PCs
HP has faced significant issues related to the high price of printer ink. The India Competition Commission has imposed fines totaling approximately 1.42 billion rupees, or around $15 million, on HP India and 21 of its resellers for manipulating bids for government contracts related to personal computers and printer supplies, including toner and cartridges. These findings pertain to tenders submitted through the Government e-Marketplace from 2017 to 2020.
Other articles
A new open-weight AI from China is surpassing the leading models from OpenAI and Claude Fable.
Moonshot's Kimi K3, with its 2.8 trillion parameters, nearly matches Fable 5 and GPT 5.6 Sol across multiple benchmarks, all while maintaining open weight.
