Tencent Hy3 and GLM 5.2 are both open-weight Chinese coding models from major tech companies, both launched in 2026, and both competing for the same developer audience. But they take different architectural bets and have different strengths. This comparison breaks down which one to pick for your work.
At a glance
| Tencent Hy3 | GLM 5.2 | |
|---|---|---|
| Vendor | Tencent (Hunyuan team) | Z.ai / Zhipu AI |
| Architecture | 295B MoE, 21B active | 744B MoE, larger active |
| Context | 256K tokens | 1M tokens |
| SWE-bench Verified | 74.4% | strong |
| SWE-bench Pro | not published | 62.1% |
| License | Apache 2.0 | MIT |
| OpenRouter price | ~$0.14/M input | varies by plan |
| Coding score | 47.37 (on par with DeepSeek V4 Pro) | competitive |
| Companion tool | none (yet) | ZCode desktop app |
| Internal deployment | 50+ Tencent products | Z.ai platform |
Where Hy3 leads
Efficiency. Hy3 achieves coding parity with much larger models using only 21B active parameters. That means lower inference cost per token and faster generation. For budget-conscious developers running at volume, this efficiency matters.
SWE-bench Verified. Hy3’s 74.4% on SWE-bench Verified is a strong result that puts it ahead of many models with far more parameters. It demonstrates that the MoE routing is doing its job well, selecting the right experts for coding tasks.
No ecosystem lock-in. Hy3 is a standalone model with weights on Hugging Face. You run it however you want: Ollama, vLLM, any compatible inference engine. There is no proprietary app or subscription plan required.
Apache 2.0 with no geographic restrictions. Previous Tencent models had license carve-outs blocking certain regions. Hy3 is clean for commercial use worldwide.
Where GLM 5.2 leads
1M context window. GLM 5.2 offers four times the context length of Hy3 (1M vs 256K). For whole-codebase reasoning, massive document processing, or long agent sessions, that gap is significant.
Integrated tooling. GLM 5.2 powers ZCode, a first-party desktop coding agent with remote control from Telegram, SSH development, and a Goal system for autonomous tasks. It also runs through Claude Code via the Z.ai compatible API. Hy3 has no equivalent app layer yet.
SWE-bench Pro. GLM 5.2 scores 62.1% on SWE-bench Pro, the harder benchmark that tests multi-file real-world issues. Hy3 has not published a Pro score, so direct comparison on this metric is not possible yet.
Established developer ecosystem. The GLM Coding Plan has been running for months with tiers from Lite to Max. Developers have built workflows around it. Hy3 is newer and lacks that infrastructure.
The architecture tradeoff
These models represent two different bets:
- Hy3: Smaller active footprint (21B), faster inference, cheaper per token. Bets on efficiency and cost.
- GLM 5.2: Larger total model (744B), wider context (1M), integrated tooling. Bets on capability ceiling and developer experience.
For coding specifically, the benchmarks suggest they are closer than the parameter counts would predict. Hy3’s 74.4% SWE-bench Verified with 21B active parameters is remarkably efficient. GLM 5.2’s 62.1% SWE-bench Pro shows strength on harder, more realistic tasks.
Practical considerations
Cost. Hy3 on OpenRouter costs approximately $0.14 per million input tokens. GLM 5.2 requires a Coding Plan subscription ($10-30/month depending on tier). For high-volume work, Hy3’s pay-per-token model is often cheaper. For consistent daily use, the GLM plan can be more predictable.
Local inference. Both can be run locally with quantization. Hy3’s 21B active parameters make it more approachable on consumer hardware (with offloading). GLM 5.2’s larger size needs more substantial hardware. See how to run Hy3 locally and our VRAM guide.
Trust and provenance. Both are Chinese companies. Both are open-weight. Both are Apache 2.0 or MIT licensed. If your compliance allows Chinese-origin models, either works. If not, neither does. See AI model supply chain risks.
Which should you choose?
Choose Hy3 when:
- Cost per token dominates your decision.
- You want the most efficient model (best results per active parameter).
- You do not need more than 256K context.
- You prefer a standalone model without a required subscription plan.
Choose GLM 5.2 when:
- You need a 1M token context window.
- You want integrated tooling (ZCode desktop agent, Claude Code compatibility).
- You prefer a subscription plan with predictable monthly cost.
- You value the established GLM developer ecosystem and documentation.
Use both when:
- You route tasks by complexity: Hy3 for routine coding at low cost, GLM 5.2 for deep, context-heavy work.
- You are evaluating which fits your workflow before committing.
For the broader Chinese model landscape, see best Chinese AI models 2026 and the Hy3 vs DeepSeek V4 comparison.
Frequently asked questions
Is Hy3 better than GLM 5.2 for coding? On SWE-bench Verified (74.4%), Hy3 posts a strong number. GLM 5.2 scores 62.1% on the harder SWE-bench Pro. They test different things, and direct comparison is incomplete without both metrics for both models.
Which is cheaper? Hy3 at ~$0.14/M tokens on OpenRouter is typically cheaper per token than a GLM Coding Plan, especially at high volume.
Which has the bigger context window? GLM 5.2 at 1M tokens versus Hy3 at 256K. Four times the difference.
Can I use both in the same workflow? Yes. Route by task: Hy3 for routine, cost-sensitive coding; GLM 5.2 for context-heavy or tool-integrated work.
Which is better for self-hosting? Hy3 is more approachable due to its smaller active parameter count (21B vs GLM 5.2’s larger footprint). Both offer open weights.
The bottom line
Hy3 wins on efficiency and cost. GLM 5.2 wins on context and tooling. For most coding work that fits in 256K tokens, Hy3 offers better value. For whole-codebase reasoning and integrated agent workflows, GLM 5.2’s ecosystem is stronger. Many teams will end up using both. Start with the Hy3 complete guide or GLM 5.2 in Claude Code depending on which matters more to you.