Inside the Mind of Tencent's Hy4: Why AI Thinks in Broken English

If you could eavesdrop on an artificial intelligence's inner thoughts, what would they sound like? Surprisingly, they don't sound like the polished, eloquent sentences we usually get on our screens. Instead, they sound a lot like a busy chef muttering in shorthand during a dinner rush.

Recently, a developer tasked Tencent’s newly released Hy4 Preview model with a whimsical prompt: "Generate an SVG of a pelican riding a bicycle." Before the AI produced the final code, it was caught having a brief, hidden internal debate.

"Let's maybe add a helmet?" the model pondered in its hidden reasoning trace. "It could improve riding theme, but may obscure head. Maybe a small cycling cap or helmet? The user didn't ask; can add red helmet? Might be cute. But pelican with big beak; a helmet might obscure. Better maybe no." It then briefly considered sunglasses before firmly deciding against them as well.

This fascinating glimpse into the machine's "mind" is a hallmark of Tencent's latest open-weight text model. Hy4 is a technological behemoth, boasting 770 billion total parameters—the virtual equivalent of brain synapses. However, it uses an efficient architecture where only 49 billion parameters are "active" at any given time, allowing it to process massive amounts of information without requiring impossible amounts of computing power. It also features a massive 1-million-token context window, meaning it can hold the equivalent of several thick books in its short-term memory.

But what makes Hy4 particularly interesting to everyday users is a behind-the-scenes setting discovered in its code called "reasoning effort." It essentially has two gears: "high" (which is the default, where it thinks out loud internally before answering) and "no_think" (where it skips the deliberation and relies on instinctual, immediate generation).

Notice the slightly broken, telegraphic grammar in the pelican debate? That’s not a bug; it's a feature of computational economics. In the world of Large Language Models, every piece of text—whether seen by the user or hidden in the background—consumes "tokens." Perfect grammar, complete with proper articles and conjunctions, requires more tokens, which in turn requires more processing time and electricity. When the AI is just talking to itself in "high" reasoning mode, it drops the pleasantries. It uses truncated English to save computational power, proving that even for machines, efficiency beats elegance when no one is watching.

As AI models continue to grow—Hy4 is more than double the size of its predecessor, Hy3, released just a month prior—the ability to toggle their "thinking" time will become crucial. We are moving away from AI that simply blurts out the most likely next word, toward systems that can pause, reflect, and debate their own choices. Understanding this hidden layer of reasoning helps demystify AI, showing us that behind the magic is a highly pragmatic, calculating assistant just trying to figure out if a pelican really needs a helmet.

Key Points

• Tencent's new Hy4 Preview model features 770 billion total parameters and a 1-million-token context window.
• A new 'reasoning effort' setting allows the model to switch between deep internal deliberation and instant responses.
• Hidden reasoning traces reveal that the AI debates its creative choices, like whether to give a pelican a helmet.
• The AI intentionally uses broken, truncated English during internal reasoning to save tokens and computational power.

Why It Matters

As AI systems become more complex, their ability to pause and reason internally represents a major leap forward. Understanding these hidden processes demystifies AI, revealing it as a pragmatic tool that constantly balances quality with computational cost.

Sources:

• Introducing Hy4 Preview — Simon Willison's Weblog