The Hy3 demand connection

The capex upgrade was not a standalone macro call — it was a direct response to the explosive adoption of Hy3. Since the Hy3 preview launch on April 23, 2026, Tencent disclosed that daily average token consumption increased 20-fold, and voluntary Hy3 preview users on WorkBuddy rose 6-fold . This usage surge created the capacity justification for the massive capex ramp. BofA's thesis hinges on the idea that demand-side proof of product-market fit — not just model capability — is what justifies the aggressive infrastructure spending.

Hy3 model specifications

Hy3 is a 295B-parameter Mixture-of-Experts (MoE) language model with hybrid fast-and-slow thinking, developed by Tencent's Hunyuan team. It activates only approximately 21B parameters per token using 192 experts with top-8 sigmoid routing . The model supports a 256K-token context window (262K reported by some sources) and can output up to 131K tokens on the full release . A 3.8B-parameter MTP (Multi-Token Prediction) speculative decoding layer accelerates inference speed .

The architecture uses 80 transformer layers (layer 0 dense, layers 1–79 MoE), Grouped Query Attention with 64 Q / 8 KV heads, per-head QK RMSNorm, and RoPE . Hy3 is released under the Apache 2.0 open-weight license, with the GA release on July 6, 2026 (preview on April 23, 2026) . Hy3 delivers flagship-level intelligence comparable to models 2–5x larger at a fraction of the compute cost, with configurable reasoning effort (no-think, low CoT, high CoT) .

Agent capabilities

Hy3 is explicitly architected for agentic workflows, not just chat or text generation :

• Hybrid reasoning – Default fast "no-think" mode plus chain-of-thought (low/high) for complex math, code, and multi-step agent tasks
• Context Learning – Excels at contextual memory and understanding for long-running agent sessions
• Agent task execution – Optimized for tool calling, code generation, search, and multi-modal agent loops
• Production-grade stability – The GA release significantly improved stability and cost-effectiveness over preview

In internal evaluations of Hy3 in WorkBuddy's office scenarios, task success rates rose from 72% to 90%, with average task duration reduced by 34% . After Yuanbao launched file delivery powered by Hy3, its common-sense error rate was cut in half, and hallucination rates dropped by more than half .

Ecosystem integrations

Hy3 is already embedded across Tencent's product suite :

• WorkBuddy – Enterprise productivity agent; BofA reports it has the largest user base of any AI agent tool in China . It underwent 43 iterations in 3 months and is widely used across Chinese enterprises .
• CodeBuddy – Full dev-cycle coding assistant covering coding, review, and deployment .
• Yuanbao – Tencent's consumer AI assistant; now supports "Lobster" integration for agentic tasks .
• ima – Personal AI agent creator with memory that learns with use; connects inside WeChat ecosystem (Docs/Meetings/Mail) .
• Marvis – IT operations AI assistant .
• Additional products – QQ, QQ Browser, Tencent Docs, Tencent Leisure, and Tencent Cloud TokenHub API . Tencent reports nearly 50 more products in the integration pipeline .

Broader AI investment ramp

Confirmed data points from evidence:

Tencent raised $4.6B in offshore bonds in June 2026 for AI development and refinancing . Q1 2026 revenue was RMB 196.5B (+9% YoY, missing consensus by ~1.5%), with the miss partly attributed to AI spending and Spring Festival timing .

Note on Goldman Sachs forecast: While Goldman Sachs and other sell-side firms have been publishing Tencent capex estimates, the exact RMB 151B figure for FY2026 often cited could not be independently verified from the available search results. One SCMP article cited a Goldman Sachs forecast of RMB 165B for 2027, but the FY2026 figure remains unconfirmed .

Strategic implications

WeChat distribution advantage

BofA's thesis explicitly hinges on Tencent's unparalleled distribution advantage: WeChat provides a built-in user base of over a billion users that can access AI agents without friction — through Yuanbao, ima, QClaw, and upcoming WeChat-native AI agents . BofA noted earlier (June 3, 2026) that a WeChat AI agent launch is a key catalyst for Tencent's re-rating .

Market view of AI agent leadership

BofA's central argument in the July 13 report is that Tencent's leadership in China's AI agent orchestration layer is underestimated by the market . The logic:

• Model → Agent → Distribution is the stack Tencent controls end-to-end. Hy3 (model) powers WorkBuddy/CodeBuddy (agents) distributed through WeChat/Yuanbao (distribution).
• WorkBuddy is #1 in China by user scale in the AI agent segment .
• The 20x token consumption growth and 6x voluntary user growth on WorkBuddy validate product-market fit and justify the capex acceleration .
• BofA expects a clear revaluation path driven by: Hy3 upgrades, capex/cloud revenue acceleration, and WeChat AI agent launch .

The capex upgrade signals that Tencent is betting the infrastructure investment will be self-funding — usage drives revenue, revenue funds more infrastructure, and the WeChat flywheel compounds the advantage. BofA's HK$750 target price implies confidence that Tencent's AI agent distribution moat, rather than just model capability, is the durable competitive advantage .

Bottom line: The Hy3 adoption surge is the demand-side proof that justifies the 23–25% capex increase — not the other way around. Tencent is now positioned as the leading end-to-end AI agent platform in China, with a distribution funnel (WeChat) that no domestic competitor can match, and the market is beginning to re-rate the stock accordingly.

The Hy3 demand connection

The capex upgrade was not a standalone macro call — it was a direct response to the explosive adoption of Hy3. Since the Hy3 preview launch on April 23, 2026, Tencent disclosed that daily average token consumption increased 20-fold, and voluntary Hy3 preview users on WorkBuddy rose 6-fold . This usage surge created the capacity justification for the massive capex ramp. BofA's thesis hinges on the idea that demand-side proof of product-market fit — not just model capability — is what justifies the aggressive infrastructure spending.

Hy3 model specifications

Hy3 is a 295B-parameter Mixture-of-Experts (MoE) language model with hybrid fast-and-slow thinking, developed by Tencent's Hunyuan team. It activates only approximately 21B parameters per token using 192 experts with top-8 sigmoid routing . The model supports a 256K-token context window (262K reported by some sources) and can output up to 131K tokens on the full release . A 3.8B-parameter MTP (Multi-Token Prediction) speculative decoding layer accelerates inference speed .

The architecture uses 80 transformer layers (layer 0 dense, layers 1–79 MoE), Grouped Query Attention with 64 Q / 8 KV heads, per-head QK RMSNorm, and RoPE . Hy3 is released under the Apache 2.0 open-weight license, with the GA release on July 6, 2026 (preview on April 23, 2026) . Hy3 delivers flagship-level intelligence comparable to models 2–5x larger at a fraction of the compute cost, with configurable reasoning effort (no-think, low CoT, high CoT) .

Agent capabilities

Hy3 is explicitly architected for agentic workflows, not just chat or text generation :

• Hybrid reasoning – Default fast "no-think" mode plus chain-of-thought (low/high) for complex math, code, and multi-step agent tasks
• Context Learning – Excels at contextual memory and understanding for long-running agent sessions
• Agent task execution – Optimized for tool calling, code generation, search, and multi-modal agent loops
• Production-grade stability – The GA release significantly improved stability and cost-effectiveness over preview

In internal evaluations of Hy3 in WorkBuddy's office scenarios, task success rates rose from 72% to 90%, with average task duration reduced by 34% . After Yuanbao launched file delivery powered by Hy3, its common-sense error rate was cut in half, and hallucination rates dropped by more than half .

Ecosystem integrations

Hy3 is already embedded across Tencent's product suite :

• WorkBuddy – Enterprise productivity agent; BofA reports it has the largest user base of any AI agent tool in China . It underwent 43 iterations in 3 months and is widely used across Chinese enterprises .
• CodeBuddy – Full dev-cycle coding assistant covering coding, review, and deployment .
• Yuanbao – Tencent's consumer AI assistant; now supports "Lobster" integration for agentic tasks .
• ima – Personal AI agent creator with memory that learns with use; connects inside WeChat ecosystem (Docs/Meetings/Mail) .
• Marvis – IT operations AI assistant .
• Additional products – QQ, QQ Browser, Tencent Docs, Tencent Leisure, and Tencent Cloud TokenHub API . Tencent reports nearly 50 more products in the integration pipeline .

Broader AI investment ramp

Confirmed data points from evidence:

Tencent raised $4.6B in offshore bonds in June 2026 for AI development and refinancing . Q1 2026 revenue was RMB 196.5B (+9% YoY, missing consensus by ~1.5%), with the miss partly attributed to AI spending and Spring Festival timing .

Note on Goldman Sachs forecast: While Goldman Sachs and other sell-side firms have been publishing Tencent capex estimates, the exact RMB 151B figure for FY2026 often cited could not be independently verified from the available search results. One SCMP article cited a Goldman Sachs forecast of RMB 165B for 2027, but the FY2026 figure remains unconfirmed .

Strategic implications

WeChat distribution advantage

BofA's thesis explicitly hinges on Tencent's unparalleled distribution advantage: WeChat provides a built-in user base of over a billion users that can access AI agents without friction — through Yuanbao, ima, QClaw, and upcoming WeChat-native AI agents . BofA noted earlier (June 3, 2026) that a WeChat AI agent launch is a key catalyst for Tencent's re-rating .

Market view of AI agent leadership

BofA's central argument in the July 13 report is that Tencent's leadership in China's AI agent orchestration layer is underestimated by the market . The logic:

• Model → Agent → Distribution is the stack Tencent controls end-to-end. Hy3 (model) powers WorkBuddy/CodeBuddy (agents) distributed through WeChat/Yuanbao (distribution).
• WorkBuddy is #1 in China by user scale in the AI agent segment .
• The 20x token consumption growth and 6x voluntary user growth on WorkBuddy validate product-market fit and justify the capex acceleration .
• BofA expects a clear revaluation path driven by: Hy3 upgrades, capex/cloud revenue acceleration, and WeChat AI agent launch .

The capex upgrade signals that Tencent is betting the infrastructure investment will be self-funding — usage drives revenue, revenue funds more infrastructure, and the WeChat flywheel compounds the advantage. BofA's HK$750 target price implies confidence that Tencent's AI agent distribution moat, rather than just model capability, is the durable competitive advantage .

Bottom line: The Hy3 adoption surge is the demand-side proof that justifies the 23–25% capex increase — not the other way around. Tencent is now positioned as the leading end-to-end AI agent platform in China, with a distribution funnel (WeChat) that no domestic competitor can match, and the market is beginning to re-rate the stock accordingly.