Tencent Hy3 vs DeepSeek V4 Pro: The New Open-Weight Price Fight
Tencent Hy3 just landed on OpenRouter, and it immediately changes the open-weight price-performance conversation. On LMRank, Hy3 enters at 8.8, just below DeepSeek V4 Pro at 9.0. That gap matters, but the pricing gap may matter more: Hy3 is listed at $0.14 per million input tokens and $0.58 per million output tokens, versus DeepSeek V4 Pro at $0.44 and $0.87.
That makes Hy3 roughly 68% cheaper on input and about 33% cheaper on output. For long-running agents, retrieval-heavy workflows, and high-volume coding pipelines, that is not a rounding error. It is the kind of delta that can move a model from “interesting benchmark result” to “default production candidate.”
Models: Tencent Hy3 · DeepSeek V4 Pro
Categories: Best Agentic Models · Best Coding Models · Best Open Models · Best Cheap Models
The headline numbers
| Model | LMRank score | Rank | Input price | Output price | Context |
|---|---|---|---|---|---|
| Tencent Hy3 | 8.8 | #24 | $0.14/M | $0.58/M | 262K |
| DeepSeek V4 Pro | 9.0 | #16 | $0.44/M | $0.87/M | 1M |
DeepSeek V4 Pro still has the stronger overall profile. It has a higher LMRank score, a full 1M-token context window, and a more established reputation across coding, STEM reasoning, and agentic benchmarks. It remains one of the best open-weight frontier alternatives for teams that want high capability without closed-model pricing.
But Tencent Hy3 is not trying to win only on raw rank. It is a 295B-parameter Mixture-of-Experts model with 21B active parameters, an Apache 2.0 license, and a 262K context window. The model page data highlights 84.2 on BrowseComp for agentic search and 79.1 on MCP-Atlas for tool orchestration. Those are exactly the areas where price compounds quickly: agents browse, call tools, retry, summarize, and consume far more tokens than a normal chat session.
Where Hy3 can beat DeepSeek V4 Pro
The case for Hy3 is strongest when workload economics dominate. At $0.14/M input tokens, Hy3 is less than one-third the input price of DeepSeek V4 Pro. If your application is retrieval-heavy, reads large contexts, or runs multi-step agent loops with frequent prompt expansion, input cost can be the majority of the bill.
Hy3 also competes well for tool-heavy workflows. Its MCP-Atlas result points at practical tool orchestration rather than only static benchmark competence. In other words: it may be a better “agent worker” model than its overall #24 rank suggests, especially if the task is structured, tool-assisted, and price-sensitive.
The Apache 2.0 license is another important difference. DeepSeek V4 Pro is open-weight and commercially permissive, but Hy3’s Apache positioning gives teams a clean, familiar legal frame for deployment and adaptation. If you are evaluating models for self-hosted or hybrid deployments, licensing simplicity can matter almost as much as benchmark deltas.
Where DeepSeek still wins
DeepSeek V4 Pro remains the safer choice when the task demands maximum capability from an open-weight model. Its 9.0 LMRank score, #16 rank, 1M-token context, and established coding and reasoning profile give it a broader ceiling. Hy3’s 262K context is long, but it is not in the same class as DeepSeek’s million-token window.
DeepSeek also has more accumulated signal. Hy3 is newer, and independent benchmark verification is still catching up. LMRank lists Hy3’s weaknesses as including text-only modality, high self-hosting GPU requirements, and pending third-party validation. That does not make it weak; it means buyers should run their own evals before replacing a known DeepSeek V4 Pro deployment.
The practical takeaway
Think of Tencent Hy3 as the new pressure point below DeepSeek V4 Pro. It does not clearly dethrone DeepSeek on absolute capability, but it can rival or beat it on price-adjusted performance for agentic search, tool use, and long-ish context production workloads.
If you need the strongest open-weight model with the deepest context, start with DeepSeek V4 Pro. If you need a cheaper agentic worker that still scores near frontier open-weight territory, test Tencent Hy3 immediately. For many production systems, the winning architecture may be both: DeepSeek V4 Pro as the high-confidence planner, Hy3 as the lower-cost execution and retrieval worker.
Sources
- OpenRouter - Tencent Hy3
- Hugging Face - tencent/Hy3
- Artificial Analysis - Hy3
- OpenRouter - DeepSeek V4 Pro
- Artificial Analysis - DeepSeek V4 Pro
Track live rankings and pricing on Tencent Hy3, DeepSeek V4 Pro, and the best open models leaderboard.