Tencent’s Hunyuan team open-sourced Hy4 preview on August 28, a 770-billion-parameter mixture-of-experts model that activates just 49 billion parameters per request. The model supports a context window exceeding one million tokens and is available through Hugging Face, OpenRouter, and Tencent’s own WorkBuddy and CodeBuddy platforms.
The release lands in the middle of an extraordinary week for Chinese open-source AI. Zhipu’s GLM-5.3-Flash was revealed as the anonymous Ox Alpha model that dominated OpenRouter. Alibaba released Qwen3.8-Flash-Next. And now Tencent has put a 770B model into the open-weight ecosystem at a price point — $0.834 per million input tokens — that undercuts most Western competitors by an order of magnitude.
The Numbers
Hy4 is a mixture-of-experts (MoE) architecture: 770 billion total parameters, with approximately 49 billion active for any given text request. This is the same architectural pattern used by DeepSeek and others to keep inference costs down while maintaining a large knowledge base. For context, Anthropic’s Claude Opus 4.8 and OpenAI’s GPT-5.6 are widely believed to use similar sparse activation strategies, though neither company has confirmed parameter counts.
Tencent conducted an internal blind evaluation involving 163 experts and 203 engineering tasks. Hy4 preview scored an average of 2.99 out of 4.00, edging out GLM-5.3 at 2.92 and Kimi K3 at 2.94. These are small margins, and the evaluation was internal — but the fact that Tencent is benchmarking against GLM-5.3 and Kimi K3, rather than Western models, tells you where the competitive frontier sits for Chinese AI labs right now.
What is a mixture-of-experts model? Instead of using every parameter for every request, an MoE model routes each input to a subset of “expert” networks within the model. This means a 770B model can run with the compute cost of a 49B model. The trade-off is memory: you still need to load all 770B parameters, even if only 49B are active at any moment.
The Self-Improvement Angle
The most novel claim in Tencent’s announcement is that Hy4 participated in its own development. The model “participated for the first time in the automated optimization of training methods, data strategies, evaluation frameworks, and low-level operators.” It proposed approaches, ran experiments, and iterated based on results — establishing what Tencent calls “an early-stage recursive self-improvement loop.”
Hy4 also autonomously analyzed bottlenecks in its inference system and carried out optimizations including operator fusion and communication optimization. The result: a 31.8 per cent increase in end-to-end throughput compared with baseline, consistent across different context lengths and concurrency levels.
This is a claim worth scrutinising. Recursive self-improvement is a term with a lot of baggage in AI safety circles, and Tencent is careful to frame it as “early-stage.” But the throughput optimization is concrete and measurable — if the gains hold in production, it’s a meaningful data point for the argument that large models can improve their own infrastructure.
What It’s Built For
Tencent is positioning Hy4 squarely at productivity tasks rather than chatbot interactions. The model is designed for coding, office work, data analysis, game development, and scientific research. Specific capabilities include:
- Software engineering: Stronger planning, debugging, and validation for long-context development tasks, plus improved front-end visual quality
- Office productivity: Cross-document collaboration, financial analysis, and full-workflow support from information processing through document creation
- Game development: Can generate a playable prototype from a single natural-language request and work with game engines
- Scientific research: Noted improvements in AI research, molecular dynamics simulation, condensed-matter physics, and fundamental mathematics
The model is available for free on WorkBuddy and CodeBuddy for two weeks following launch. API access costs $0.834 per million input tokens, $2.501 per million output tokens, and $0.042 per million tokens for cache hits.
The Honest Caveats
Tencent is upfront about limitations. The preview “may take too long when working through complex questions” and “may over-verify its own answers.” For research and financial analysis use cases — exactly the domains Tencent is targeting — a model that spends excessive time reasoning and double-checking could be a product friction point rather than a feature.
Superpower Daily noted that this caveat is “particularly relevant to the research and financial-analysis uses Tencent names, where a lengthy response process can become part of the product experience.” In other words, a model that takes 30 seconds to answer a financial question might be thorough — or might be unusable in a real-time workflow.
The Open-Source Arms Race
Hy4’s release continues a pattern that’s now unmistakable: Chinese AI labs are using open-source releases as a competitive weapon. As we’ve argued before, this isn’t charity — it’s a distribution strategy. Every developer who downloads Hy4, GLM-5.3, or Qwen is one fewer developer locked into a Western API. The pricing makes the play explicit: at $0.834 per million input tokens, Hy4 costs roughly one-fortieth of what Anthropic charges for Opus 4.8 at comparable intelligence levels, according to Zhipu’s own comparison.
The difference is that Tencent’s model is explicitly aimed at enterprise productivity — coding, analysis, document creation — rather than the frontier coding and cyber capabilities that GLM-5.3 emphasises. It’s a complementary play, not a head-to-head competition with Zhipu.
FAQ
Is Hy4 free? The model weights are open-sourced on Hugging Face. WorkBuddy and CodeBuddy offer free access for two weeks. API access is paid at $0.834/M input, $2.501/M output tokens.
How big is 770 billion parameters? For comparison, Meta’s Llama 4 has roughly 400B total parameters. Hy4 is nearly double that, though only 49B activate per request.
Can I run it locally? In theory, yes — the weights are open. In practice, loading 770B parameters requires roughly 1.5 terabytes of VRAM, which means multi-GPU server setups. This is a data centre model, not a laptop model.
How does it compare to GPT-5.6 or Claude? Tencent’s internal evaluation compared Hy4 against GLM-5.3 and Kimi K3, not Western models. No independent benchmark comparison against GPT-5.6 or Claude Opus 4.8 has been published yet.
