Tencent Hunyuan Drops 'Hy3' Official Model: Massive MoE Upgrade Challenges DeepSeek V4 Pro and Triggers New API Price WarFollowing months of intense market anticipation surrounding its high-performing preview releases, the Tencent Hunyuan research team has officially launched the production-ready iteration of its flagship AI model: Tencent Hy3. The preview version had already captured significant market share among developers due to ultra-aggressive third-party hosting prices sitting at just $0.06 per million input tokens and $0.21 per million output tokens. The finalized production model delivers substantial architecture upgrades, narrowing the performance gap to rival massive proprietary models like DeepSeek V4 Pro and GLM-5.2.
Tencent’s engineering telemetry highlights a massive leap forward in reasoning stability and reliability. In specialized autonomous coding benchmarks across leading IDE assistants such as Tencent’s native CodeBuddy, as well as popular third-party tools like Cline and KiloCode the production model dropped its tool-use and agentic execution error rates to a mere 4%. Furthermore, in standard multi-turn conversational benchmarks, the model’s hallucination rate plummeted from 12.5% in the preview branch down to 5.4% in the official release.
The model weights have been officially uploaded to HuggingFace, revealing a massive 295B-A21B Mixture-of-Experts (MoE) architecture packaged in BF16 precision (meaning a total of 295 billion parameters, with 21 billion parameters activated per token). In mainland China, Tencent has pegged the official first-party API pricing at 1 RMB per million input tokens and 4 RMB per million output tokens. While this represents roughly double the cost of the heavily subsidized third-party Hy3 Preview instances currently floating on the market, it remains an existential threat to competitors clocking in at three times cheaper on input costs compared to DeepSeek.
For global developers utilizing OpenRouter, NovitaAI is currently offering free access to Hy3 for a two-week promotional window, while AtlasCloud has spun up paid production endpoints priced symmetrically to Tencent’s official domestic rates.
Tencent Hy3 Production Model Overview
The Developer Group: Tencent Hunyuan Research Team.
The Architecture Framework: Mixture-of-Experts (MoE) with a 295B total / A21B activated parameter configuration.
The Precision Format: Brain Floating Point 16-bit (BF16).
The Open-Weights Repository: Available for local deployment via HuggingFace.
The Reliability Metrics: Coding tool error rate dropped to 4%; general hallucination rate slashed from 12.5% to 5.4%.
The Price Disruptor: Official first-party API priced at 1 RMB/4 RMB per million tokens (~$0.14/$0.55). Input costs are 3x cheaper than DeepSeek.
Global Access Gateways: Free for 2 weeks on OpenRouter via NovitaAI; commercially hosted via AtlasCloud.
The parameter code 295B-A21B represents a Mixture-of-Experts (MoE) architecture. While the model boasts a massive 295 billion parameters for extensive knowledge storage, comparable to DeepSeek V4 Pro, during runtime it only activates 21 billion parameters relevant to the question. This design avoids burdening the model with computational tasks, resulting in extremely high throughput and significantly reduced power and server costs. This allows Tencent to aggressively undercut the market price.
This "API Price Under-cutting" strategy, following DeepSeek's shockwaves in Silicon Valley with its inexpensive model (priced at just a few cents), is a counter-strategy employed by Tencent, a major player with a larger cloud presence and deeper financial resources. By pricing input tokens three times cheaper than DeepSeek, Tencent deliberately targeted developers of AI agents and large-scale enterprise chatbots. These systems primarily rely on inputting lengthy context windows. The three-fold reduction in input costs could save startups millions of dollars per month—a price point that DeepSeek itself might find difficult to match if it were to cut its prices to compete.
Tencent's demonstration of Hy3's stability on computational tools like Cline, KiloCode, and CodeBuddy is another key aspect. In today's AI world, answering general questions in an entertaining way or composing poetry is no longer the sole measure of intelligence; rather, it's the ability to "function calling" or "tool use"—such as instructing the AI to open files, read code, debug, and execute terminal commands instead of humans. Hy3's ability to reduce the error rate when working with these tools to just 4% is a game-changing metric, proving that this open-source model is ready for use in large-scale enterprise software industries, no longer just a toy for the general public.
Qualcomm Linux 2.0 Goes Live Migrating to GitHub to Challenge Raspberry Pi IoT Reign.
Source: Tencent
Tencent Hunyuan Drops 'Hy3' Official Model: Massive MoE Upgrade Challenges DeepSeek V4 Pro and Triggers New API Price WarFollowing months of intense market anticipation surrounding its high-performing preview releases, the Tencent Hunyuan research team has officially launched the production-ready iteration of its flagship AI model: Tencent Hy3. The preview version had already captured significant market share among developers due to ultra-aggressive third-party hosting prices sitting at just $0.06 per million input tokens and $0.21 per million output tokens. The finalized production model delivers substantial architecture upgrades, narrowing the performance gap to rival massive proprietary models like DeepSeek V4 Pro and GLM-5.2.
Tencent’s engineering telemetry highlights a massive leap forward in reasoning stability and reliability. In specialized autonomous coding benchmarks across leading IDE assistants such as Tencent’s native CodeBuddy, as well as popular third-party tools like Cline and KiloCode the production model dropped its tool-use and agentic execution error rates to a mere 4%. Furthermore, in standard multi-turn conversational benchmarks, the model’s hallucination rate plummeted from 12.5% in the preview branch down to 5.4% in the official release.
The model weights have been officially uploaded to HuggingFace, revealing a massive 295B-A21B Mixture-of-Experts (MoE) architecture packaged in BF16 precision (meaning a total of 295 billion parameters, with 21 billion parameters activated per token). In mainland China, Tencent has pegged the official first-party API pricing at 1 RMB per million input tokens and 4 RMB per million output tokens. While this represents roughly double the cost of the heavily subsidized third-party Hy3 Preview instances currently floating on the market, it remains an existential threat to competitors clocking in at three times cheaper on input costs compared to DeepSeek.
For global developers utilizing OpenRouter, NovitaAI is currently offering free access to Hy3 for a two-week promotional window, while AtlasCloud has spun up paid production endpoints priced symmetrically to Tencent’s official domestic rates.
Tencent Hy3 Production Model Overview
The Developer Group: Tencent Hunyuan Research Team.
The Architecture Framework: Mixture-of-Experts (MoE) with a 295B total / A21B activated parameter configuration.
The Precision Format: Brain Floating Point 16-bit (BF16).
The Open-Weights Repository: Available for local deployment via HuggingFace.
The Reliability Metrics: Coding tool error rate dropped to 4%; general hallucination rate slashed from 12.5% to 5.4%.
The Price Disruptor: Official first-party API priced at 1 RMB/4 RMB per million tokens (~$0.14/$0.55). Input costs are 3x cheaper than DeepSeek.
Global Access Gateways: Free for 2 weeks on OpenRouter via NovitaAI; commercially hosted via AtlasCloud.
The parameter code 295B-A21B represents a Mixture-of-Experts (MoE) architecture. While the model boasts a massive 295 billion parameters for extensive knowledge storage, comparable to DeepSeek V4 Pro, during runtime it only activates 21 billion parameters relevant to the question. This design avoids burdening the model with computational tasks, resulting in extremely high throughput and significantly reduced power and server costs. This allows Tencent to aggressively undercut the market price.
This "API Price Under-cutting" strategy, following DeepSeek's shockwaves in Silicon Valley with its inexpensive model (priced at just a few cents), is a counter-strategy employed by Tencent, a major player with a larger cloud presence and deeper financial resources. By pricing input tokens three times cheaper than DeepSeek, Tencent deliberately targeted developers of AI agents and large-scale enterprise chatbots. These systems primarily rely on inputting lengthy context windows. The three-fold reduction in input costs could save startups millions of dollars per month—a price point that DeepSeek itself might find difficult to match if it were to cut its prices to compete.
Tencent's demonstration of Hy3's stability on computational tools like Cline, KiloCode, and CodeBuddy is another key aspect. In today's AI world, answering general questions in an entertaining way or composing poetry is no longer the sole measure of intelligence; rather, it's the ability to "function calling" or "tool use"—such as instructing the AI to open files, read code, debug, and execute terminal commands instead of humans. Hy3's ability to reduce the error rate when working with these tools to just 4% is a game-changing metric, proving that this open-source model is ready for use in large-scale enterprise software industries, no longer just a toy for the general public.
Qualcomm Linux 2.0 Goes Live Migrating to GitHub to Challenge Raspberry Pi IoT Reign.
Source: Tencent
Comments
Post a Comment