Tencent released Hy3, the official version of its Hunyuan model, on July 8. The headline: a 295-billion-parameter model with only 21 billion activated parameters outperforms models two to five times its size on agentic and reasoning tasks—and it's open-source under Apache 2.0.

What Hy3 Ships With

The model specs are concrete:

  • 295B total parameters, 21B activated — a mixture-of-experts (MoE) architecture that activates only a fraction of weights per token, reducing latency and memory footprint.
  • 256K context window — enough for multi-step agent reasoning over moderately long documents.
  • Benchmarked against: GLM-5.1 (744B total, 40B activated), DeepSeek V4 Pro, Qwen 3.7 Max, and GLM-5.2. In Tencent's published evaluations, Hy3 sits between GLM-5.1 and GLM-5.2 on agent tasks; comparable to DeepSeek V4 Pro on reasoning.
  • Compared to Hy3 Preview (April 2026), the official release gains in code generation, agent capabilities, and stability. Tencent says it has been tested on internal office-agent and productivity workflows.

Why This Matters for Agents

Agent workloads are not traditional chat. A coding agent, office assistant, or tool-using model must chain reasoning, instruction following, long-context understanding, code generation, tool calls, and error recovery in a single workflow. Raw parameter count tells you almost nothing about whether a model can do that reliably.

Hy3's positioning is explicit: compete on capability per active parameter, not sheer scale. That's the right move for a world where inference cost and latency matter more than leaderboard rankings. Smaller, well-optimized models are easier to:

  • Run on-premise or self-hosted without $100M/month cloud bills.
  • Cache and reuse (Tencent Cloud pricing includes cache hits at 0.25 yuan per million tokens, versus 1 yuan for input).
  • Integrate into agents that need sub-second response times.
  • Iterate on locally without paying per-token API costs.

Pricing and Distribution

On Tencent Cloud, the API pricing is:

  • 1 yuan per million input tokens
  • 4 yuan per million output tokens
  • 0.25 yuan per million cache-hit input tokens (useful for agents that reuse context)

For high-volume agentic use (long-context reasoning, repeated queries over the same documents), caching is material.

Hy3 will be available on Tencent Cloud TokenHub, Hugging Face, ModelScope, GitHub, and global platforms: OpenRouter, Hermes, Kilo, Cline, OpenClaw, OpenCode, CherryStudio. That distribution strategy matters—adoption isn't decided by benchmarks alone. It's decided by whether a developer can download the weights, integrate it into their existing tooling, and move fast.

The Catch: Vendor Benchmarks

As the article notes, Tencent's comparisons will need independent testing. Vendor-published benchmarks are notorious for optimistic framing. The real test is whether independent developers, agent builders, and AI reviewers can reproduce those results in their own agentic workflows—coding tasks, tool use, multi-step reasoning, error recovery—in the wild.

The timeline is also worth watching: Hy3 Preview shipped in April, the official version in July. A two-month iteration loop on a foundation model is fast for Tencent, and the company is using that pace to signal that Hunyuan is moving from leaderboard optimization to product feedback. If that holds, we may see Hy3 improves based on real-world agent usage rather than benchmark tuning.

What This Signals

As AI moves deeper into coding assistants, office agents, and enterprise automation, the market is likely to care less about flagship scale and more about practical efficiency: models powerful enough to build with, affordable enough to run at volume, and open enough to adopt without licensing friction. Hy3 is Tencent's bet that smaller, well-optimized open models can own that space.

Sources: