Tencent's Hunyuan team officially released Hy3 on July 12 — a hybrid fast/slow thinking MoE model with 295B total parameters and only 21B active, a 256K context window, and a fully permissive Apache 2.0 license (available on HuggingFace and ModelScope).
Key specs
- Architecture: MoE with fast/slow thinking mode switching (same lineage as Hunyuan-A13B), significantly scaled-up post-training relative to the Hy3 preview
- Context: 256K tokens, covering full codebases of mid-sized projects
- License: Apache 2.0, free for commercial use worldwide
- Pricing: 1 RMB per million input tokens, 4 RMB per million output tokens via Tencent Cloud TokenHub; overseas API platforms to follow
- Task success rate: 72% (preview) → 90% (official release)
- Average completion time: down 34%
- Daily token consumption of Hy3 preview grew 20x; users choosing it autonomously grew 6x
- Task completion rate: 93.7%, up 12.7 points from preview
- Hy3: 2.67/4 vs. GLM-5.1: 2.51/4
- Leading categories: frontend, data & storage, CI/CD
- Hallucination rate: 12.5% → 5.4%
- Common-sense error rate: 25.4% → 12.7%
- Known weaknesses: pure vision (video understanding, complex chart parsing); 256K context trails 1M-window rivals like Claude Code / GPT-5.6 Sol; structured-output/tool-call reliability needs third-party validation
- Data credibility: 72%→90% success rate is official Tencent data; independent blind tests pending. The distribution of improvements (easy vs. hard tasks) matters more than the average
- Sustainability: Yuanbao Agent's "all free" strategy and the feasibility of "queueing for value" under high concurrency remain to be seen; independent HuggingFace/LMSYS data expected in August
- Tencent official announcement: https://www.tencent.com/zh-cn/articles/2202386.html
- Tencent Cloud developer community analysis: https://cloud.tencent.com/developer/article/2707339
- NetEase Tech report (270-expert blind test data): https://www.163.com/dy/article/L168CHBE0514R9OJ.html
- STCN "capacity crunch" scoop: https://new.qq.com/rain/a/20260710A04K9N00
- AYi AI Notes hands-on video: https://x.com/AYi_AInotes/status/2076341952023310580
Production benchmarks (not leaderboard scores)
WorkBuddy agent tasks (automated script generation + workflow orchestration):
Marvis Agent (file editing, management, PC diagnostics):
270-expert internal blind test:
Capacity crunch (July 8): WorkBuddy compute was overwhelmed with queue rates over 50%; Tencent expanded capacity overnight and restored service by the morning of July 9.
Positioning: an "Agent-oriented LLM"
Hy3's core thesis, per chief researcher Yao Shunyu: "In the Agent era, sufficient capability + low cost + thorough open source wins developers." Rather than pursuing uniform SOTA across all dimensions (pure vision, hard math proofs), Tencent invested its optimization budget in the engineering line — cost, latency, reliability, debuggability — matching the needs of its product matrix (WeChat, Yuanbao, WorkBuddy, Marvis, CodeBuddy, ima, WeChat Reading, WeGame, WeGame game assistant).
| Vendor | Flagship | Route | |--------|----------|-------| | DeepSeek | V4 series | Ultra-large MoE + inference optimization | | Qwen | 3.7 | Dense + MoE dual versions | | GLM | 5.2 | Dense model + code specialization | | Meta | Llama 4 | Ultra-large MoE | | Tencent Hy3 | 295B/21B MoE | MoE + product-driven + Apache open source |
Industry context (2026 H2)
1. Chinese open-source models breaking the price-performance ceiling — alongside DeepSeek V4 Flash RL training on AMD Instinct MI355X, Meituan's LongCat-2.0 on a 50K domestic GPU cluster, and Xiaomi's MiMo-7B coding release 2. Agent evaluation shifting from benchmarks to production data — ByteDance's EdgeBench, Senior SWE-Bench, Ploy's redesign suite all move toward real work scenarios 3. Domestic AI coding toolchain acceleration — Tencent's stack (CodeBuddy + Hy3 + Yuanbao Agent + Marvis), Alibaba's Page Agent + Qwen, ByteDance's Doubao + Coze, Meituan's LongCat + Tabbit