“Fable-level immersive experience” 2.8 trillion-parameter Kimi K3 sets a new open-source record

Author: Li Dan, Wall Street Insights

China’s next round of “open-source wars” for large models is officially underway as the Dark Side of the Moon (“Moon’s Dark Side”) launches.

On Thursday, July 16, the Dark Side of the Moon officially released its new-generation open-source foundation model Kimi K3, with a parameter scale of 2.8 trillion, and simultaneously rolled out API services and developer documentation. Xinhua reported that in evaluations, Kimi K3’s overall intelligence was close to leading closed-source models worldwide, and it said this is the open-source model with the largest number of parameters globally at present, marking another step forward in the development of artificial intelligence (AI) models in China.

Kimi K3 is the strongest flagship model from the Dark Side of the Moon to date, built specifically for long-range agent programming and self-evolving workflows. It introduces the company’s self-developed Kimi Delta Attention (KDA) hybrid linear attention mechanism and the Attention Residuals (AttnRes) structure, improving long-sequence information processing and deep-network information transfer by optimizing training efficiency and inference performance for large-scale models.

New technology enables up to 6.3x decoding acceleration for one million Token context

Kimi K3 supports a 1 million Token context window and natively has visual understanding capability. It is mainly aimed at high-difficulty scenarios such as software engineering, knowledge work, and complex reasoning. The Dark Side of the Moon says that KDA achieves up to 6.3x decoding acceleration under million-level context, while the AttnRes mechanism boosts training efficiency by about 25% with additional costs of less than 2%.

The Dark Side of the Moon emphasizes that this model focuses on optimizing agentic coding capabilities, enabling it to understand large codebases, call tools, execute tests, and continuously adjust task plans based on feedback. At the same time, K3 uses a higher-sparsity MoE design. According to information compiled by the community, the model has 896 expert modules, but only 16 experts are activated per inference; this expands model capacity while controlling computational costs.

Compared with the previous generation, Kimi K3 places greater emphasis on shifting from “answering questions” to “completing tasks.” Especially in long-cycle software engineering tasks, the model can make integrated judgments by combining code, runtime logs, test results, and visual information. This gives it stronger potential in scenarios such as game development, front-end engineering, CAD design, and infrastructure optimization.

Multiple tests approach top closed-source models, with some evaluations exceeding Claude Opus 4.8

After the release of Kimi K3, performance testing became a key focus for the market.

Based on the evaluation results published by the Dark Side of the Moon and data compiled by the community, Kimi K3 has moved into the global first-tier across multiple areas including code generation, knowledge work, long-text retrieval, and agent tasks. In some comprehensive capability evaluations, K3 is seen as ranking just behind the Claude Fable series and the top GPT models, placing among the leading models in the test set.

In knowledge-work evaluations, community-compiled GDPval-AA v2 data shows that Kimi K3 scored 1687 points, surpassing Claude Opus 4.8 Max’s 1600 and ranking only behind Claude Fable 5 Max and GPT-5.6 Sol Max. This test mainly evaluates the real work ability of the model across 44 job categories and 9 industries, covering complex tasks such as research analysis, business judgment, and professional writing.

In addition, in the AA-Briefcase agent knowledge-work test, Kimi K3 scored 1527 points, ranking only behind Claude Fable 5 Max and surpassing GPT-5.6 Sol Max.

In terms of long-context and information retrieval capabilities, community data shows that Kimi K3 scored 91.2 points in the BrowseComp test. Because the model supports a 1 million Token context window and can complete tasks in a single-agent mode without additional context compression, it is considered to have clear advantages in long-cycle, high-difficulty information retrieval scenarios.

“Fable-level feel” sparks discussion as Kimi K3 challenges Anthropic’s closed-source route

Beyond Benchmark results, Kimi K3’s real-world experience in the developer community has also sparked a lot of discussion.

Some early testing users said that in complex agent tasks, continuous coding, and multi-tool calling scenarios, K3’s autonomous planning and execution abilities are close to the level previously achievable only by top closed-source models. Some developers have called it having “Fable-level feel.”

One evaluation result mentioned by an overseas hedge fund manager is that Kimi K3’s performance exceeds Opus 4.8, and its pricing is about 60% of Opus 4.8’s. An in-China bank team’s testing also concluded that Kimi K3’s capabilities exceed Opus 4.8 and are close to Fable 5 and GPT-5.6 Sol.

However, compared with Anthropic’s Claude Fable series, Kimi K3’s more accurate positioning at present is “approaching,” not “fully surpassing.”

Because models like Fable have not released full parameter details and all test specifics, outsiders cannot conduct strict one-to-one comparisons. But based on published evaluations, Kimi K3 has already surpassed Claude Opus 4.8 in some knowledge-work and agent capability tests, showing that China’s open-source models are entering a new phase of competing with the U.S.’ top closed-source models.

Industry insiders believe the significance of Kimi K3 is not just its parameter scale and Benchmark scores, but also its choice to open up weights. Unlike OpenAI and Anthropic, which mainly provide closed-source models through APIs, the Dark Side of the Moon aims to expand the developer ecosystem through open-source models.

After DeepSeek, Kimi K3 further demonstrates that Chinese AI companies are shifting from competing on cost advantages to competing comprehensively on model capabilities, agent ecosystems, and developer infrastructure.

Overseas attention before release: benchmark to Anthropic’s flagship model

In fact, days before Kimi K3 officially debuted, overseas media had already begun closely watching the model.

Reports said that inside the company, Kimi K3 is positioned as one of the largest AI models currently in China. It is expected to outperform Anthropic’s Claude Opus 4.8 in multiple mainstream benchmark tests and further narrow the performance gap with Anthropic’s flagship model Claude Opus 4.8. This is said to be the most important model upgrade from the Dark Side of the Moon to date.

Although Anthropic has not disclosed the parameter scale of Opus 4.8, industry estimates generally put it at around 1.5 trillion to 2 trillion parameters.

The Financial Times, citing people familiar with the matter, said that while Kimi K3 is still expected not to surpass Anthropic’s Fable, the earlier-advanced model that was paused due to safety issues, it is already enough to challenge the common belief in the market that “Chinese models are 8 to 12 months behind the U.S.”

For U.S. AI labs, the bigger challenge is not only model capability, but the business model.

Because Kimi K3 is released with open weights, developers worldwide can download it for free, deploy it themselves, and modify it. This means it has the potential—like DeepSeek—to quickly form a developer ecosystem, creating ongoing competitive pressure on OpenAI and Anthropic, which stick to closed-model strategies.

AI competition focus: not just performance, but cost and open ecosystem

Over the past year, U.S. AI companies have continued to invest thousands of billions of dollars to build AI infrastructure and have kept rolling out more powerful frontier models.

Meanwhile, their commercialization pricing has also continued to rise.

According to information on Anthropic’s official website, starting this September, the company plans to further increase the price of Claude Opus 4.8 by about 50%, raising the input Token price to $3 per million Tokens and the output Token price to $15 per million Tokens.

By contrast, Chinese AI companies are taking a different path.

Including DeepSeek and the Dark Side of the Moon, multiple Chinese AI companies have continued to roll out open-weight models. These not only allow enterprises to deploy privately, but also keep overall inference costs significantly lower than those of top U.S. models.

Taking the Dark Side of the Moon’s previously released K2.6 model as an example, its calling cost is about one-third of Claude Opus 4.8’s.

As more companies focus on AI cost control, many overseas firms have also started trying Chinese models to replace part of their U.S. models in order to reduce inference spending.

Marc Andreessen, a co-founder of Andreessen Horowitz, a well-known U.S. venture capital firm, previously said that Zhipu’s GLM-5.2 has already become the first Chinese model that can match— and in some public tests even surpass—U.S. large AI lab flagship models.

Open-source competition keeps heating up as China’s AI company value is reappraised

Since this year began, nearly all Chinese AI companies have shifted to open-source routes.

DeepSeek quickly drew global developer attention with its R1 series; companies such as Zhipu and MiniMax then followed with open-weight models; and now the Dark Side of the Moon has officially joined this camp.

At the same time, top U.S. AI companies still stick to closed-source routes.

OpenAI’s latest GPT series and Anthropic’s flagship Claude models have not opened model weights. Instead, they rely on an API subscription model to keep boosting commercial revenue.

This difference in model approach has become a new dividing line in global AI competition.

The Financial Times points out that an increasing number of Silicon Valley investors and technology company executives have started to believe that the performance gap between the frontier AI models of China and the U.S. is shrinking quickly, and that what will truly determine the future competitive landscape may no longer be just model rankings, but who can build an open ecosystem that covers developers worldwide.

The capital markets are also reassessing the value of Chinese AI companies.

The Financial Times reported that the Dark Side of the Moon is conducting a new round of financing with a valuation of about $31.5 billion; DeepSeek has also started a new round of financing with a valuation of about $71 billion. By comparison, Anthropic’s valuation after its latest funding is about $965 billion, and OpenAI’s latest valuation is about $852 billion.

With Kimi K3 officially released, competition between China’s leading AI companies and U.S. frontier labs is also extending beyond a pure model-capability contest, moving further into comprehensive battles over open-source ecosystems, developer communities, and business models.