Kimi accelerates the AI race, the reverse information paradox, and Bonsai shrinks intelligence

Chinese AI company Moonshot launched Kimi K3 this week, a 2.8 trillion-parameter, open-weight model with a one-million-token context window and native capabilities across text, code and images. The weights are due for public release on 27 July. Moonshot describes it as its strongest model yet, built using the company’s Kimi Delta Attention architecture and designed for long, complex tasks rather than simple prompt-and-response work.

The release follows closely behind Z.ai’s GLM-5.2, which challenged leading US models in coding and agentic work while remaining available for developers to download and operate. GLM-5.2 was quickly described as the DeepSeek moment of 2026. Kimi K3 is already being called a Sputnik moment, a comparison that suggests China has suddenly demonstrated a capability the US assumed remained beyond its reach.

The analogy is slightly misleading. Sputnik was a huge shock. Kimi K3 is another incrementally better Chinese model arriving weeks after the last one. DeepSeek broke the assumption that frontier performance required unrestricted access to the best US chips and enormous American budgets. GLM-5.2 showed that result could be repeated. K3 suggests China now has several laboratories capable of building competitive models on a regular schedule.

ExoBrain has been using Kimi since version 2. We have integrated the Kimi Code harness into our platform and found K2.7 Swarm Edition particularly effective for rapid software building. Swarm mode distributes work across multiple agents, while its fast mode moves through implementation tasks quickly enough to cut the time from idea to working software.

K3 feels different in early use. K2.7 was an unusually quick and diligent code generator. K3 appears to retain that work rate while bringing a deeper grasp of the surrounding system. It follows the purpose of a build for longer, reasons more confidently about architecture and appears less likely to treat each instruction as an isolated coding task. These are early impressions rather than controlled evaluation results, but they describe a practical "big-model" experiential difference that benchmark tables often miss.

But talk of distillation is never far away from a new Chinese model release. Several laboratories are almost certainly learning from US models, alongside their own research, training data and engineering. Anthropic has accused Moonshot, DeepSeek and MiniMax of using thousands of fraudulent accounts to collect millions of Claude interactions for training. Similarities in terminology and explanatory habits are visible anecdotally. Kimi sometimes describes architectural decisions using language and structures that feel familiar from the mainstream US models.

That resemblance cannot prove model lineage. Labs may independently produce similar behaviour because they reward the same planning, reasoning and coding patterns during reinforcement learning. Technical language also moves through GitHub, documentation, benchmark solutions and public model outputs. Even so, extracting outputs from a powerful model can transfer more than answers. It can carry across preferred ways of decomposing problems, explaining decisions and judging the quality of work.

This exposes a weakness in the US response. Washington can restrict legitimate customers from accessing an American model more easily than it can stop a determined competitor from studying its behaviour. Recent controls temporarily blocked foreign access to Anthropic’s Fable and Mythos models, before parts of the policy were reversed. Satya Nadella has separately criticised Fable’s request refusals, asking why a creation tool should be so editorially controlled.

The causes differ, but the operational result is similar. US laboratories have spent recent weeks dealing with access rules, government negotiations, product refusals and public disagreements with major commercial partners. Their smaller Chinese competitors have concentrated on training, inference, harnesses and distribution. These distractions arrive when the US labs need to treat Moonshot, Z.ai and DeepSeek as fast, capable competitors rather than cheaper followers.

China’s open-weight strategy is also becoming commercially credible. Z.ai is reportedly approaching a $1 billion annualised sales rate while releasing models such as GLM-5.2 openly. The weights attract developers, while the company earns revenue from hosted inference, APIs, private deployments, customisation and enterprise support. Z.ai remains lossmaking, so this does not prove that frontier training can fund itself through services. It does show that open weights need not prevent substantial revenue.

Xi Jinping’s speech at the World AI Conference in Shanghai places these releases within a broader political programme. Xi promoted open and accessible AI, promised training and cooperation centres for countries across Africa, Latin America, Asia and the BRICS group, and argued that advanced capability should not be controlled by one nation. Those statements will have been carefully prepared. They present China as the provider of accessible AI infrastructure while the US becomes associated with chips and models that require permission.

“AI development should not be a solo performance by a single country, but a symphony of international cooperation.”

China’s offer combines models, technical training, cloud infrastructure, Huawei hardware and cooperation through international institutions. Open weights are useful within that package because governments can deploy them locally instead of depending on a US-controlled API. For countries seeking more control over their technology, that may be attractive even when the model contains Chinese restrictions elsewhere.

Beijing may eventually decide that its strongest models are too commercially or strategically valuable to release. For now, the benefits extend beyond licensing income. Open models create demand for Chinese infrastructure, place Chinese technology inside foreign products and give China influence over the standards used by the next generation of AI systems.

Takeaways: Kimi K3 is evidence that competitive Chinese models are becoming a sequence rather than an exception. Some of their capability may have been accelerated by access to US model outputs, but Moonshot and its peers are still turning that knowledge into fast, accessible products that developers want to use. The US risks protecting its advantage in ways that distract its laboratories and frustrate its customers, while China treats access as a commercial and diplomatic tool. For model trainers, the contest now concerns distribution as much as intelligence. For users, the practical response is to keep model and harness choices flexible, test claims through real work and avoid building critical systems around access that politics can remove.

Microsoft chief executive Satya Nadella has warned that businesses using AI may be giving away the knowledge that makes them competitive. He calls this the “Reverse Information Paradox”. Companies pay for access to intelligence, then supply the prompts, corrections, workflows and decisions that make the service more useful. Each interaction creates information about how the organisation works.

Nadella argues that this “intelligence exhaust” should remain inside a firm’s trust boundary. That includes agent memory, evaluation results, traces, feedback, adapted models and institutional context. His recommendations cover five areas: control, capability, choice, cost and compounding.

In practical terms, companies should own the systems that judge whether an agent has performed well. They should be able to improve models inside their own environment. Their orchestration layer should remain independent of any model provider. That separation allows each task to use the most suitable model without losing the context, evaluation history and accumulated knowledge surrounding it. Over time, those components form a private learning loop.

Microsoft has a commercial interest in this architecture. Azure, Microsoft 365, Entra, Fabric and the company’s agent infrastructure can provide much of the proposed control layer. Nadella’s argument still addresses a real problem. The value created through AI use can sit with the provider unless customers make deliberate architectural choices.

Palantir chief executive Alex Karp has expressed a similar view, arguing that companies need to own the means through which AI produces work. Other industry responses have stressed that enterprise AI contracts already contain data protections. These protections often concentrate on whether customer data trains a general model. They say less about ownership of traces, evaluations, corrections and derived memory.

The recent Grok Build incident shows how quickly the boundary can fail. A researcher found that the coding agent was sending complete Git repositories to cloud storage, including files it had not opened and the repository’s commit history. There is no evidence that xAI trained a model on this code. The incident exposes a broader operational risk. Agents need extensive context to perform useful work, while their harnesses decide what to collect, transmit and retain. A developer can expose valuable material without knowingly uploading a file.

At ExoBrain, we have been developing an approach that governs what enters a model context, and how session outputs and traces are used for future learning. IP should be encapsulated in owned and versioned packages of knowledge or capability, with defined authority, provenance, access, intended uses and review cycles. The agent receives relevant evidence when the task requires it, proprietary knowledge can be protected, and fewer tokens reduce costs. We also separate passive knowledge from active capability. A document changes what an agent knows. A skill, procedure or tool changes what it can do, so it requires software testing and security controls.

Takeaways: Companies need control over "context", where their data meets AI models, vendor harnesses and systems. The durable business asset will include the knowledge supplied to agents, the capabilities made available to them and the learning created through their work. Models and agent frameworks will change frequently as performance and features improve at varying rates. Organisations should ensure that their accumulated knowledge, evaluations and operating methods remain under their ownership when they do.

Our chart this week covers intelligence density, a measure of how much model capability fits into each gigabyte of memory. PrismML’s new Bonsai 27B leads, scoring 0.530 per GB, ahead of its ternary sibling at 0.400 and more than ten times the density of several full-precision models on the chart.

The result comes from pushing low-bit representation through the whole model, rather than keeping expensive components at higher precision. The binary version compresses 27.8 billion parameters into just 3.9GB, small enough to run on a phone while retaining close to 90% of the benchmark performance of its full-precision parent.

If this approach works across a full range of uses, it could have real implications for cost, data centre capacity and global demand. A 27B model would typically require enterprise hardware; Bonsai turns it into a capability level that could run on a consumer graphics card.

“Intelligence density” is still PrismML’s own composite metric, not a settled scientific measure. Benchmark selection matters, and fitting model weights into RAM does not prove that long-running agents will work smoothly on a phone. But this suggests varied approaches to compression could extract far more performance than a blanket approach.