On July 16, a Beijing startup most people outside tech circles had never heard of released a 2.8 trillion parameter language model and put the entire file online for anyone to download. No paywall, no waitlist, no enterprise sales call. Just the weights, free to take.

That startup is Moonshot AI, and the model is Kimi K3. It is now the largest open-weight AI system ever released, and its arrival says less about one company’s ambition than about a strategic fork between the US and China over how artificial intelligence should be distributed.

The launch that rattled a market

Moonshot did not undersell it. The company said K3 had reached what it called “open frontier intelligence”, a phrase built to plant a flag rather than describe a benchmark.

At 2.8 trillion parameters, the model is roughly 75 per cent larger than DeepSeek’s V4 Pro and dwarfs Zhipu AI’s 744 billion parameter GLM-5 series.

Parameters are not everything, and Moonshot was upfront that K3’s overall performance still trails the most powerful proprietary models.

It did nevertheless claim K3 was “consistently outperforming other tested models”, including OpenAI’s GPT-5.5 and Anthropic’s Claude Opus 4.8 across a range of evaluations.

On specific tests such as Program Bench and SWE Marathon, Moonshot says K3 even edged out newer systems including GPT-5.6 Sol and Claude Fable 5. Those are the company’s own reported figures, and independent verification is still catching up.

Investors did not wait for confirmation. Shares tied to rival labs Z.ai and MiniMax fell 27 per cent and 16 per cent respectively in the hours after the announcement, according to reporting compiled by MLQ News.

Interest in the model was so intense that Moonshot suspended new subscriptions within 48 hours of launch after demand outpaced its compute capacity, according to the Global Times.

The full weights were not even available when the initial market reaction hit. Researchers who reviewed Moonshot’s technical documentation told VentureBeat that the complete files were scheduled to drop on July 27, more than a week after the announcement.

Two engineering tricks, and a pricing shot across the bow

Underneath the headline parameter count, Moonshot is leaning on two homegrown architectural ideas: Kimi Delta Attention, a hybrid linear attention mechanism, and Attention Residuals.

The company describes Attention Residuals as a replacement for standard residual connections designed to keep performance scaling smoothly as the model grows. Both ideas were published as open research on GitHub before K3 shipped.

The model also carries a 1 million token context window and what Moonshot calls an always-on “thinking mode”, designed for large codebases or long documents in a single pass.

On efficiency, the company says K3 needs 21% fewer output tokens than its predecessor, K2.6, to complete equivalent tasks.

Then there is price. Moonshot set API access at $3 per million input tokens and $15 per million output tokens. MLQ News described that as the highest rate among Chinese labs, while still putting the per-task cost at roughly half that of Anthropic’s Opus 4.8.

Moonshot needed the money too. The company raised $2 billion at a $20 billion valuation in May and is reportedly now in talks to raise another $30 billion.

The real story is the strategy

Strip away the benchmark wars and a clearer pattern emerges. American labs including OpenAI, Anthropic and Google are largely keeping their frontier models behind APIs, while guarding architecture details and even basic facts such as parameter counts.

Forbes noted that neither OpenAI nor Anthropic has disclosed the total parameter count of their top models. Moonshot’s decision to publish the number, along with the weights, is therefore a pointed contrast.

Chinese labs have gone the other way. DeepSeek did it. Alibaba’s Qwen line did it. Zhipu’s GLM did it. Moonshot has now done it at a scale nobody else has attempted.

The Global Times quoted Pan Helin, a member of the Information and Communication Economy Expert Committee under China’s Ministry of Industry and Information Technology, describing the move as deliberate positioning.

Moonshot was, he said, “open-sourcing the infrastructure technologies” partly to answer doubts about its own algorithmic work.

That is the strategic bet in a sentence. Rather than compete purely on raw capability, where US labs still hold an edge according to many independent evaluations, Chinese developers are competing on access.

Give away the weights, undercut the price, and let anyone from a startup in Lagos to a university lab in São Paulo build directly on the technology instead of a closed American system.

This is not Moonshot’s first swing

Moonshot was not a lock to be a major player this year. Its market position had eroded sharply over the previous 18 months, largely because DeepSeek’s sudden rise reshaped China’s AI landscape in early 2026.

K3 marks a genuine comeback, but it builds on an approach Moonshot had already been testing with earlier releases.

When Moonshot released Kimi K2, an earlier 1 trillion parameter model with 32 billion active parameters, the reception among developers was already enthusiastic.

Perplexity CEO Aravind Srinivas wrote at the time that “Kimi models are looking good on internal evals”, adding that his team would likely begin post-training on it soon.

That kind of unsolicited endorsement from a Western AI company is precisely the credibility Moonshot is trying to accumulate again with K3, only at a much larger scale.

Reuters reported on the K2 launch as part of a broader wave, noting that Moonshot said the release outperformed mainstream open-source models in some areas, including DeepSeek’s V3, while rivalled leading US models on specific functions such as coding.

Why this matters beyond the leaderboard

The practical consequence for users is a widening of choice. A business in Southeast Asia or Eastern Europe that cannot justify enterprise pricing from Anthropic or OpenAI can now download, self-host and modify a frontier-adjacent model.

That is a real shift in who gets to build serious AI products, rather than simply who gets access to a chatbot. Open weights can reduce dependence on a single provider and give developers greater control over deployment.

There is a catch that often gets lost in the excitement. A 2.8 trillion parameter model is not something most companies can run on their own hardware.

VentureBeat pointed out that the model’s sheer size could make private deployment impractical for many businesses across the Asia-Pacific region. Many users may still have to rent access through cloud infrastructure.

That weakens some of the independence that open weights are supposed to provide. The software may be freely available, but the computing power needed to run it remains expensive.

There is also the question of trust. Kimi, like the other models discussed here, comes from a China-based company. Organisations handling regulated or sensitive data will have to assess that risk before routing information through it.

The bottom line

Kimi K3 probably is not the smartest model in the world, and Moonshot has never claimed that it is. What it represents is something more structural: a different strategy for competing in AI.

Chinese developers are increasingly betting that the fastest way to close the gap with the US is not to out-build American labs behind closed doors, but to out-distribute them in the open.

Whether that strategy works will depend less on this week’s benchmark charts than on how many developers in how many countries decide it is easier to build on a free download than a metered API.