Tencent Hy3 (Hunyuan 3.0)
Tencent's open-weight 295B MoE reasoning/agent model — its first frontier-class release, led by ex-OpenAI researcher Shunyu Yao.
1. Core Product / Service
Hy3 preview is the third generation of Tencent's Hunyuan foundation-model line and its first release positioned as frontier-class. It is a Mixture-of-Experts model with 295B total parameters and 21B activated (192 experts, top-8 routing, plus a 3.8B-parameter MTP layer for speculative decoding), with a 256K-token context window [1][2]. Weights were open-sourced on Hugging Face, ModelScope, GitHub, and GitCode in late April 2026 under the Tencent Hy Community License (Tencent's press release dates the announcement April 24, 2026; the GitHub repo states April 23) [1][2].
The release is explicitly agent-first: Tencent claims it "reliably powers complex agent workflows of up to 495 steps" and emphasizes MCP toolchain orchestration, document processing, and search. Day-one inference support landed in vLLM (MTP speculative decoding) and SGLang (EAGLE), both with OpenAI-compatible APIs [1][2]. Commercial API access runs on Tencent Cloud at $0.18 / $0.59 per million input/output tokens, with a limited two-week free window on openrouter at launch [1].
2. Target Users & Pain Points
Two distinct audiences: Tencent's own consumer/productivity ecosystem (the model is already live inside Yuanbao, QQ, QQ Browser, Tencent Docs, ima, and the CodeBuddy/WorkBuddy coding assistants, where Tencent reports a 54% TTFT reduction and 47% faster end-to-end responses) [1], and external developers running agentic workloads who want frontier-adjacent capability at small-active-parameter cost. The 21B-active MoE design targets the same pain point deepseek established: serving cost scales with active parameters, so a 295B/A21B model can undercut dense rivals of similar capability on inference price.
3. Competitive Landscape
| Model (2026) | Architecture | Open weights | SWE-bench Verified |
|---|---|---|---|
| Hy3 preview | 295B MoE / A21B | Yes (community license) | 74.4% [3] |
| Claude Opus 4.6 (Anthropic) | undisclosed | No | 80.8% [3] |
| GLM-5 (zhipu) | MoE | Yes | 77.8% [3] |
| Kimi K2.5 (kimi) | MoE | Yes | 76.8% [3] |
| DeepSeek V3.x (deepseek) | MoE / sparse | Yes | — |
| Qwen 3.x (qwen) | dense + MoE zoo | Yes | — |
Hy3 jumped from 53% (Hy2) to 74.4% on SWE-bench Verified — a 40% generational improvement that puts Tencent in range of the leading Chinese open-weight labs after years of trailing them, though still slightly behind GLM-5 and Kimi K2.5 on that benchmark [3]. Its differentiation is cost-efficiency (fewest active parameters in its capability class) and the distribution moat of Tencent's app ecosystem rather than raw benchmark leadership.
4. Unique Observations
- Hy3 marks Tencent's strategy shift from "embrace DeepSeek" to "compete with DeepSeek." The tencent-cloud page documents how Tencent fully integrated deepseek into Yuanbao in 2025 when Hunyuan 2 lagged; shipping a competitive in-house open-weight model reverses that dependency while keeping the WeChat/QQ distribution advantage.
- It is the first major output of Tencent's Shunyu Yao era — a bet that hiring a single high-profile agent-research lead (ReAct, SWE-bench lineage at OpenAI) can compress the gap with deepseek/kimi/zhipu faster than scaling internal teams. The agent-centric framing (495-step workflows, MCP orchestration) reflects that pedigree.
- The 295B/A21B shape extends the trend Jimmy tracks in ai-token-supply-chain: Chinese labs competing on tokens-per-dollar rather than peak capability, pushing L2 (model) margin compression onto inference providers. A frontier-adjacent model with only 21B active parameters at $0.18/M input is another step down the price curve set by DeepSeek.
5. Financials / Funding
Not a standalone company — funded internally by Tencent. Disclosed pricing: $0.18 per million input tokens / $0.59 per million output tokens on Tencent Cloud TokenHub [1]. No separate revenue figures for the Hunyuan line are disclosed.
6. People & Relationships
- Key person: Shunyu Yao — Chief AI Scientist at Tencent; former OpenAI core researcher, recruited to lead Tencent's AI effort; Hy3 is the first major release under his leadership [1][3].
- Parent: Tencent (distribution via Yuanbao, QQ, Tencent Docs; commercial API via tencent-cloud).
- Competitors: deepseek, qwen, kimi, zhipu, MiniMax — the Chinese open-weight frontier cohort.
- Distribution partners: openrouter (launch free window), vLLM and SGLang (day-one engine support) [1][2].
Sources
[1] Tencent, "Tencent Unveils Hy3 preview; Model Enhances Agent Capabilities and Real-World Usability," https://www.tencent.com/en-us/articles/2202320.html (2026-06-12) [2] GitHub, "Tencent-Hunyuan/Hy3-preview," https://github.com/Tencent-Hunyuan/Hy3-preview (2026-06-12) [3] Decrypt, "Tencent's New Hy3 AI Model Is the Most Efficient Chinese LLM No One's Talking About," https://decrypt.co/365297/tencent-hy3-preview-open-source-moe-model (2026-06-12)
Local sources:
2026-06-07-summary.md— Hy3 Preview flagged for tracking among new model releases