Tencent Hy Research Unveils Hy4 Model: Outperforming Kimi K3.1 Pro at a Fraction of the API CostFollowing months of public preview testing, the Tencent Hy research team has officially launched its flagship Hy4 foundation model. The production release demonstrates massive performance leaps across multimodal benchmarks outperforming industry heavyweights like the 3-trillion-parameter Kimi K3.1 Pro across several key evaluation suites while drastically undercutting market pricing on API token consumption.
Technical Advancements, Specialized Vision Training, and API Economics
Multimodal Benchmarks and Specialized Spatial Vision:
Benchmark Breakthroughs: Hy4 registered substantial score increases compared to its early preview iterations across standardized evaluations, including ProgramBench, 3DArena, and Terminal Bench 4.0.
Spatial and Graphical Vision Training: Tencent achieved these gains by implementing specialized vision training datasets. The model was trained extensively on complex 3D modeling interfaces, software environments like Blender and CAD platforms, as well as full-stack web and game development workflows.
Intentional Capability Guardrails and Acknowledged Limitations:
Safety Alignment Adjustments: Tencent explicitly restricted specific offensive capabilities within Hy4. Safety tuning deliberately dialed back penetration testing protocols, vulnerability exploitation techniques, reverse-engineering logic, and advanced bioweapons/biological synthesis data below baseline regulatory thresholds.
Identified Areas for Improvement: The engineering team transparently acknowledged specific areas where Hy4 requires further optimization, including advanced multi-step complex mathematics, long-horizon continuous coding tasks, autonomous AI model development workflows, and ultra-high-density complex image parsing. These remain key priorities for future model iterations.
Cost Architecture and Upcoming Hy4 Flash Variant:
Aggressive API Pricing: Standard commercial API pricing for Hy4 is set at $0.834 per million input tokens and $2.50 per million output tokens.
Promotional Discount: Tencent is offering a 50% promotional discount across all API tiers during the initial launch phase.
Hy4 Flash Roadmap: The research team confirmed active development of Hy4 Flash, a lightweight variant optimized for low latency, higher throughput, and even lower inference operational costs.
Source: Tencent Hy
Tencent Hy Research Unveils Hy4 Model: Outperforming Kimi K3.1 Pro at a Fraction of the API CostFollowing months of public preview testing, the Tencent Hy research team has officially launched its flagship Hy4 foundation model. The production release demonstrates massive performance leaps across multimodal benchmarks outperforming industry heavyweights like the 3-trillion-parameter Kimi K3.1 Pro across several key evaluation suites while drastically undercutting market pricing on API token consumption.
Technical Advancements, Specialized Vision Training, and API Economics
Multimodal Benchmarks and Specialized Spatial Vision:
Benchmark Breakthroughs: Hy4 registered substantial score increases compared to its early preview iterations across standardized evaluations, including ProgramBench, 3DArena, and Terminal Bench 4.0.
Spatial and Graphical Vision Training: Tencent achieved these gains by implementing specialized vision training datasets. The model was trained extensively on complex 3D modeling interfaces, software environments like Blender and CAD platforms, as well as full-stack web and game development workflows.
Intentional Capability Guardrails and Acknowledged Limitations:
Safety Alignment Adjustments: Tencent explicitly restricted specific offensive capabilities within Hy4. Safety tuning deliberately dialed back penetration testing protocols, vulnerability exploitation techniques, reverse-engineering logic, and advanced bioweapons/biological synthesis data below baseline regulatory thresholds.
Identified Areas for Improvement: The engineering team transparently acknowledged specific areas where Hy4 requires further optimization, including advanced multi-step complex mathematics, long-horizon continuous coding tasks, autonomous AI model development workflows, and ultra-high-density complex image parsing. These remain key priorities for future model iterations.
Cost Architecture and Upcoming Hy4 Flash Variant:
Aggressive API Pricing: Standard commercial API pricing for Hy4 is set at $0.834 per million input tokens and $2.50 per million output tokens.
Promotional Discount: Tencent is offering a 50% promotional discount across all API tiers during the initial launch phase.
Hy4 Flash Roadmap: The research team confirmed active development of Hy4 Flash, a lightweight variant optimized for low latency, higher throughput, and even lower inference operational costs.
Source: Tencent Hy
Comments
Post a Comment