The new country of free intelligence speaks Mandarin.

On 16 July 2026, at the World Artificial Intelligence Conference in Shanghai, the Chinese company Moonshot AI unveiled Kimi K3: a model of 2.8 trillion parameters, with a context window of one million tokens and a dated promise to publish the full weights eleven days later^[2][3]. The weights appeared on Hugging Face on 27 July, and with them the largest open model in existence: 2.8 trillion total parameters with 104 billion active per token, under the Kimi K3 License, which authorizes modifying it, continuing its training and redistributing it with attribution, and demands a separate agreement only from whoever resells it as a service above twenty million dollars^[1].

Silicon Valley and Washington read the news with a single word, Sputnik, which is worth taking apart slowly. What happened that week was not that one country humiliated another, but that frontier capability stopped being something you buy and became something you download. For any region that is not in that race — and Latin America is not — that change of nature is the whole story; K3 is merely the largest example of a phenomenon that has been accumulating for three years.

What was published

K3 is a mixture-of-experts architecture that activates 16 of its 896 experts per token, so its cost scales with the fraction switched on and not with the full 2.8 trillion. Moonshot added two innovations of its own that together speed up decoding roughly 6.3 times and improve per-token efficiency by 25%^[2].

The numbers that moved the market were the leaderboard's, not the size's. On GDPval-AA v2, a benchmark built on real tasks from 44 occupations and 9 industries, K3 scored 1,687 points and came third in the world, behind only Claude Fable 5 Max (1,815) and GPT-5.6 Sol Max (1,747.8)^[2], and it topped the frontend code arena as well^[5]. Independent evaluators warn that those tables are in large part self-reported and that the model spends reasoning generously^[4]: cheap per token is not cheap per answer. The published price is US$ 3 per million input tokens and US$ 15 for output^[3], against the roughly US$ 50 of output charged by the leading American models^[6].

The gap that closed cost hundreds of billions

A gift is measured by what it costs to produce. The four American giants have committed around US$ 725 billion in capital investment for 2026, 77% more than the US$ 410 billion of 2025^[10]. On that mountain was built the advantage K3 put within reach of anyone with an internet connection, and a Beijing lab reached third place without matching that budget. One analyst put it as a question: what exactly are the American frontier labs spending their money on?^[5]

The semiconductor slide was read as confirmation, and there the chronology dismantles the story. The sell-off was already eight days old when K3 was announced: by 8 July the Philadelphia Semiconductor index was down 10.8%, Intel more than 20%, and some US$ 1.3 trillion had been wiped from the sector over doubts about the return on AI capex, without any Chinese model figuring in the diagnosis^[7]. The two companies that know the real demand for silicon, because they manufacture it, did the opposite of retreating. On 15 July ASML raised its revenue forecast to €43–45 billion alongside 30% more EUV capacity, and on 16 July TSMC lifted its annual guidance with capex of US$ 60–64 billion^[8]. Whoever holds the hard data on demand bought more factory, not less, so what survives the noise is not that compute has stopped mattering but that leadership in models has stopped being a defensible moat.

The key weighs 1.4 terabytes

The free model breaks against physics the moment anyone tries to switch it on, because 2.8 trillion parameters in MXFP4 — four bits, quantized during training itself — take up exactly 1.4 terabytes on disk alone^[9]. That calls for some eighteen 80 GB accelerators just to load it, and of the latest generation: it runs natively on Nvidia's Blackwell and AMD's MI400^[9]. The model is free and the key is not, so for almost everyone using K3 will still mean renting it through an API: the dependency does not disappear, it changes creditor.

The West opened that door once and left it shut

The record deserves fairness, because the United States published open weights too: in August 2025 OpenAI released gpt-oss under Apache 2.0, and it remains its main open distribution with no committed cadence^[11]. Meta traveled the opposite road, and faster, because in late 2025 it shelved Llama's successor and pivoted to a proprietary line that publishes no weights^[12][13]. The result is a reversal of roles: the sustained publication of frontier capability has become Chinese practice, while the ecosystem that invented modern free software treats it as an exception already discharged.

Intelligence arrives compressed in a language that is not ours

There is a technical reason, not a cultural one, why a model does better when it is spoken to in English or in Chinese. A model is a compression of its corpus, and training corpora are dominated by English to the point that tokenizers are optimized for it^[14]. The same technical text consumes around 1.55 times more tokens in Spanish than in English with the GPT-4 tokenizer and some 1.24 times with the newer ones, while Llama 3.2 compresses 4.9 characters per token in English against 3.6–3.8 in Spanish and in Portuguese^[15].

That difference is paid three times over: in money, because every answer consumes more tokens; in memory, because a fixed context holds less of our text than of someone else's; and in quality, because the worst-compressed language is the worst-represented one. The regional benchmarks confirm it from the other end: Trueque and Choclo, from CENIA, found that no leading model exceeds 0.48 out of 1.0 in Latin American cultural knowledge, neither the Chinese one nor the American^[19].

The region has an answer, and it arrived late

Latin America had presented its own five months earlier, on 10 February 2026. Latam-GPT is serious work: Chile's CENIA coordinated it together with more than sixty institutions and two hundred specialists from fifteen countries^[18]. Its spec sheet deserves the same frankness. Latam-GPT is Llama 3.1 at 70 billion parameters, with continued pretraining on some 297 billion regional tokens, under the Llama 3.1 Community License and trained on AWS^[16][17]. The architecture of that sovereignty rests on the weights of an American company, under that same company's license and on another company's cloud, just as the company that donated the base got out of the business of donating bases.

The calendar, however, weighs more than the ownership. The base the region chose in February 2026 had been published in July 2024^[16], and its own base model still had no publication date^[17] when K3 appeared five months later with 2.8 trillion parameters, third place in the world and a price of zero. The distance between the two is not one generation but two years on a curve measured in months, and in that arithmetic a sovereign bet that takes two years to deliver is born obsolete: the capability pursued as the frontier in February is downloaded free in July. The public decision cycle does not run at the speed of the thing it is trying to acquire.

That lag explains the tone in which the region received K3. It got to the news quickly and told it well, with CNN en Español and the Argentine, Mexican and Colombian technology press covering it within days^[24]. The framing, however, was that of spectacle: the Chinese AI shaking Silicon Valley as the headline, and the price per token as a secondary detail. It is accurate coverage, and a spectator's, in which the one question the region has to decide for itself barely surfaces.

Those are the facts. What they say about which layer is worth buying is the business of the section that follows.

For three years the Latin American conversation about sovereignty in artificial intelligence was a conversation about models of our own, and that question has just been settled by donation. Nobody in the region is going to train 2.8 trillion parameters from scratch, and the reason is not a lack of talent but of financial scale: the figure to match is the US$ 725 billion the four giants committed for a single year^[10]. Starting from weights that others published is not a shortcut, then, but the only door there is. Dependency is inevitable, and what does get chosen is which one, and how many of its layers you own.

That is where the difference lies between a borrowed base and a foundation. Latam-GPT did the right thing with the hand it held, and its problem is not the fine-tune but the publisher of the base: it built on weights whose owner abandoned the practice of publishing them^[12][13], so the house ended up on land that stopped being leased. The way out exists and almost nobody in the region has named it: the base can be changed. Applying to a frontier open model the recipe that was applied to Llama — continued pretraining with the regional corpus, tuning, evaluation — is viable today on a foundation that is published on a cadence and whose license expressly authorizes it^[1], so that decision ought to be taken by looking at who publishes and under what license, before any table of benchmarks.

Changing base demands, however, the layer nobody financed: compute of our own in the region, on iron that is controlled here. The numbers placed side by side are brutal, because switching on a single model asks for 1.4 terabytes and eighteen accelerators^[9] against a Chilean supercomputer costing five million dollars and a Brazilian plan of R$ 23 billion that is still a plan^[20]. The political ceiling aggravates the arithmetic, since the region was left on the second rung of the American diffusion rule, the rule was repealed without a stable replacement, and the exemptions in force expire at the end of 2026^[22]. Access to the iron is a concession open to revision and not an acquired right.

The region holds one card the rest do not, and it is clean watts with somewhere to put them. Brazil concentrates 37.3% of the region's data centers on an electricity mix that is 88.2% renewable and exempts equipment imports from duty under the REDATA regime, while Querétaro grows faster than any other market on the continent^[21]. It is the lesson of the last edition seen from the other end: the infrastructure of artificial intelligence is an energy problem before it is a silicon one, and energy is the one thing in this stack that is not imported. Trading green megawatts and land for compute capacity installed and governed locally is a negotiation the region can win; renting tokens is not.

There remains the cheapest and most neglected layer, where the advantage is immediate. The corpus and the evaluation do not require eighteen accelerators, nobody can donate them and no global model is going to solve them for us: the tokenization tax makes every answer in Spanish between 24% and 55% more expensive^[15], and no leading model reaches half of the regional cultural bar^[19]. That 0.48 out of 1.0 is not a complaint but a product specification, and whoever defines the exam ends up defining the product that gets built to pass it. Brasília, meanwhile, legislates risk with PL 2338^[23], which is necessary and insufficient: regulating a stack you rent is consumer protection and not sovereignty.

THE PLAY

1. Choose the base by its publisher, not by its leaderboard. Weights are a commodity input and their price is going to zero, so what has to be evaluated is who publishes, under what license and at what cadence. A fine-tune on an abandoned base is an asset with an expiry date.
2. Negotiate watts for installed compute, not tokens for a credit card. The regional advantage is a clean grid with land available: trade it for locally governed capacity, with contracts that survive a change in export rules.
3. Take ownership of the corpus and of the exam. It is the only layer that depends neither on eighteen accelerators nor on a donor's goodwill, and today it is measured and vacant: 0.48 out of 1.0 is at once the diagnosis and the product plan.