Moonshot AI’s Kimi K2 outperforms GPT-4 in key benchmarks — and it’s free

Moonshot aiThe beginning of the popularity of Chinese artificial intelligence Who is the chatotChallenges a direct source language model on Friday to direct ownership systems Open and Anthropical especially with strong performance in coding and autonomous agent positions.

New model, called As k2The architecture of experts has 1 trillion general parameters with a parameter of 32 billion. The company releases two versions: an optional model for researchers and developers an optimized option for conversations and autonomous agent applications.

? Hello, like K2! Open Source Agentic Model!
? 1t Total / 32B Active Moe Model
? Sota on the Sota bench, TAU2 & Acebench between open models
? Coding and agent are strong in positions
? Multimodal and Thought mode is not supported now

Like K2 advanced agent intelligence … pic.twitter.com/plrqnrg9jl

– Like.ai (@kimi_moonshot) 11 July 2025

“As K2 does not respond; it moves,” he said Ad Blog. “Like K2, the advanced agent intellect is more open and accessible than ever. We cannot wait for you to see what you build.”

The model of the model is the optimization of the “agent” capabilities – to use tools, write and execute the code and execute the ink very step-staged tasks. In assessment tests, As k2 65.8% achieved the accuracy Swe-bucketA difficult program engineering benchmark prefers the most open source alternatives and fits some property models.

David met with Goliath: K2 K2, how to models of silicone valley models of the billion dollars K2

Performance measurements explaining a story that has to do managers Open and Anthropical Pay attention. As K2-INSTRUCT Not only competing with the great players – systematically prefers them to most enterprise customers.

Side LivecodebenchUndoubtedly, the most realistic coding criterion, As k2 53.7% accuracy, decisively beating Deepseek-v3‘s 46.9% and GPT-4.1‘s 44.7%. More Surprisingly: 97.4% hit Math-500 Compared to 92.4% of GPT-4.1, Moonshot has cracked something fundamental on mathematical justification deviating larger, better-funded opponents.

However, this is what benchmarks do not seize: Moonshot These results achieve these results with a model that costs these results to a faction spent on training and conclusion. Openai is calculated to growing hundreds of millions of millions of millions, the Moonshot seems to find a more efficient way to the same destination. This is a classic innovative game that plays in real time – scrapy edge, not only compatible with the current performance, better, faster and cheaper.

The effects are simply stretched out of the rights that are bragging. Enterprise clients were waiting for the AI systems that could autonomy complex work without creating really effective demos. The power of K2 as Swe-bucket Finally suggests that he can surrender in that promise.

Muonclip Bog: Why this Optimizer can change the AI training economy

It is a detail that is buried in the technical documents of Moonshot, which is more important than the model’s benchmark scores: their development Muonclip OptimizerFixed training “Zero training instability”, which is stable of the trillion parameter model.

This is not just an engineering achievement – potentially paradigm turn. The involvement of the training, secret taxes related to the development of the large language model, forced the expensive exercises, forced expensive training and accepting their suboptimal performance to avoid accidents. Moonshot’s solution is noticeable by resetting weight matrices directly in the survey and basic predictions by solving the problem than the use of the Lent-AIDS below.

Economic effects are amazing. Mood Muonclip proves generalized – and Moonshot Offers – Techniques can dramatically reduce the placement of large models. In an area where the training costs are measured in ten millionths, even modest effectiveness is translated into competitive advantages measured in a quarter, not the year.

More interesting, it represents a major disagreement in the philosophy of optimization. When Western AI laboratories are combined in the variability of AdamW, the optimization view of Moonshot’s moon options shows that they are exploring true different mathematical approaches. Sometimes the most important innovations are due to scaling existing methods, but completely interrogate basic assumptions.

Open Source such as Competitive Weapons: Moonshot’s radical price strategy target Big Tech’s profit centers

Moonshot’s decision to open As k2 At the same time, when offering a competitive priced API entry, it reveals a complex concept of market dynamics that go beyond the altruistic open source principles.

For cache hits a million signs and $ 2,15 for $ 2,15 for a million dollars Moonshot aggressively price Open and Anthropical Although they do not compare – and in some cases superior – performance. But the actual strategic master master is the double availability

This creates a trap for current providers. If they meet the price of moonshot, they squeeze their edges on those with the most affordable product line. If they do not, a model of customer flaws in a model carrying part of the cost. Meanwhile, Moonshot accepts market share and ecosystem through both channels at the same time.

Open source component is not charity – this is to get a customer. Each developer who uploads and practices with As k2 A potential enterprise becomes a client. Each developer that contributes by the society reduces its development costs of Moonshot. When building an impossible competitive supports to repeat competitors for accepting competitors, the global developer is a videos that accelerate the global developer community.

From Demo to Reality: As soon as K2’s agent’s capabilities signal the endbot theater

Demonstrations Moonshot Shared social media, reveals something more important than effective technical capabilities – finally indicates the AI by completing the practical aid program of salon recommendations.

See the example of salary analysis: As k2 Only did not answer the questions about the data, performed 16 python operations to create statistical analysis and interactive visualization. London participates in many platforms, searches, calendars, emails, flights, habitats and restaurant orders within the concert planning demonstration. They do not make demos designed to make a impression; The complexity of knowledge employees is samples from the AI systems that complement the multi-step work flows.

This represents a philosophical turn from the current generation of AI assistants who are superior to the conversation, but the execution. When the opponents are focused on making more people sounding more people, Moonshot made them more useful. Distinction is important, because the Enterprises do not need AI that can pass the TURING TEST – AI that can pass the productivity test.

The true progress is not in any capacity, but also in the seamless orchestra of the service. Previous attempts “Agent” in the AI, the extensive urban engineering, careful work flow and demanded permanent human control. As k2 Task splitting, tool choice and error recovery, appears to manage the cognitive order of the difference between a thin calculator and the original thinking style.

Great convergence: Open source models finally caught the leaders

As the release of K2, the forecast of industrial observers, but records an infection that rarely witnesses: open source AI opportunities are really combined with the property alternatives.

Unlike the “GPT killers”, which are superior to narrow domains, as in practical applications, K2 demonstrates extensive powers along the full spectrum of tasks that determine general exploration. The code writes, solves maths, uses tools and complements complex work flows – freely available for modification and self-placement.

This approach is especially sensitive to AI employees. Openai faces pressure pressure to justify it $ 300 billion evaluation Anthropically, increasingly struggling to distinguish the clum in a crowded market. Both companies set up pre-designed work models in the protection of technological advantages that K2 could be Ephemeral.

Timing is not accidental. Transformer architecture Mature and training techniques are increasingly varies from democracy, competitive advantage to deployment efficiency, expenditure optimization and ecosystem effects. Moonshot This passage easily understands, as it is better like a better chatBot, like K2,, place as a more practical foundation for the next generation of AI applications.

Now the question is that open source models are not able to match the owners and prove them already. The question is that the benefits of the current persons will not be able to compare the preferences of the main technologies in a world, which is no longer defenseless. Based on the release of Friday, the adaptation period is significantly short.