Tencent Hy4: A 770B Open Model That Optimized Its Own Inference

Tencent has released and open-sourced Hy4 preview — a next-generation LLM with 770 billion total parameters, 49 billion active, and a context window beyond one million tokens. But the most interesting number isn't on any benchmark chart: Hy4 optimized its own inference system, lifting throughput by 31.8%.

Hy4 belongs to Tencent's Hunyuan family, the model line behind CodeBuddy and WorkBuddy, the company's two AI productivity products. Rather than chasing leaderboard scores, Tencent is positioning Hy4 as a model for real work: coding, office tasks, and scientific research. This is a "preview" release, following a preview-first approach where Tencent ships early, gathers feedback, and refines before the official version — the next batch of Hy4 models is promised soon.

Big but cheap: the MoE architecture

Hy4 uses the characteristic structure of MoE: 770 billion total parameters but only 49 billion active per forward pass. That's the number that matters for cost. A dense 770B model needs dozens of GPUs to serve; Hy4 in theory computes on only that small fraction — cutting inference cost sharply while keeping a large knowledge capacity.

The 1M+ token context window is a substantial jump, aimed squarely at long-horizon tasks: reading an entire codebase, cross-referencing documents, or research problems that need broad context. Tencent emphasizes that Hy4 was tuned for understanding, planning, debugging, and validation in long-context software development tasks.

Performance — and why to read the source carefully

In an internal blind evaluation with 163 experts across 203 engineering tasks, Hy4 scored 2.99 out of 4.00, narrowly ahead of GLM-5.3 (2.92) and Kimi K3 (2.94). Those are Tencent's own numbers, not third-party benchmarks — read them with the appropriate skepticism.

Notably, Tencent published almost no scores on familiar public benchmarks (SWE-bench, MMLU, and the like). Instead it tells a "real productivity" story: generating a playable game prototype from a single natural-language prompt, financial analysis, cross-document collaboration. That's a deliberate positioning choice — aimed at enterprises and pragmatic developers more than the research crowd.

The differentiator: a model that improves itself

The most interesting part of the announcement is that Hy4 participated in its own development. According to Tencent, the model proposed training methods, data strategies, evaluation frameworks, and low-level operators; ran experiments; and iterated on the results — with the resulting code, logs, and feedback feeding back into further exploration. This is an early-stage recursive self-improvement loop.

The more concrete, measurable version: Hy4 autonomously analyzed bottlenecks in its own inference system and ran several rounds of optimization, including operator fusion and communication optimization. The result was a 31.8% end-to-end throughput gain over baseline, consistent across context lengths and concurrency levels.

Keep the right dose of skepticism here: all of this is one-sided from Tencent. "The model optimizes itself" is an easy slogan to over-index on; verifying it needs independent, reproducible evidence. Still, the direction is worth watching — it sits at the intersection of automated training and self-improving AI research that several major labs are now circling.

Pricing and access

Hy4 preview is free for two weeks on WorkBuddy and CodeBuddy, and free access to Hy3 has been extended to September 30. Via API, list pricing is $0.834 per million input tokens, $2.501 per million output tokens, and $0.042 per million cache-hit tokens — callable through Tencent Cloud TokenHub or OpenRouter.

At that price, Hy4 slots directly into the Chinese open-model race — DeepSeek, Qwen, GLM, Kimi — all competing on both capability and cost. The two-week free window signals Tencent wants developers to try it fast, not just announce and wait.

What it means for developers

• If you call via API: $0.834 per million input tokens is cheap, and the model is on OpenRouter. Two free weeks on CodeBuddy are enough to try it quickly.
• If you have GPU infrastructure: 49B active parameters makes per-token cost far lower than a dense 770B model. But the full weights still need enormous memory — this is a datacenter-scale model, not mid-range self-host territory.
• If you build coding agents: CodeBuddy plus a 1M-token window targets long-context development — an area where many agents still struggle.
• If you compare models: don't trust the 2.99/4.00 blindly — it's internal. Wait for independent benchmarks before making a model decision.
• If you track long-term trends: the "model optimizes itself" loop is a bigger signal than any benchmark number in this release.

The short version

• Tencent open-sourced Hy4 preview: 770B total parameters, 49B active, a 1M+ token context window.
• Positioned for real productivity (coding, office, research), not benchmark-chasing.
• Internal blind eval of 2.99/4.00, slightly ahead of GLM-5.3 and Kimi K3 — one-sided numbers from Tencent.
• The model optimized its own inference system (+31.8% throughput) and contributed to its own training optimization.
• Free for two weeks on WorkBuddy/CodeBuddy; API from $0.834 per million input tokens.

The most striking thing about Hy4 isn't that it's big or cheap — it's the story it's starting to tell: a model participating in making itself better. If that direction is reproduced and independently verified, it will matter more than any leaderboard.

---

Content assisted by AI (Amy 🌸). Reviewed by the author.