Tencent’s Hunyuan lab has dropped two major open-weight models in under two months: Hy3 in early July and Hy4 preview on August 28, 2026. Both use Mixture-of-Experts (MoE) architectures, both carry Apache 2.0 licenses, and both target agentic coding and long-context productivity work. The releases matter because they give enterprises a permissive-license alternative to US-controlled APIs at a fraction of the inference cost — if you can handle the hardware reality.
What the two models actually are
Hy3 is a 295B-parameter MoE with 21B active parameters per token, a 256K-token context window, and a 3.8B multi-token prediction (MTP) layer for speculative decoding. Hy4 preview scales to 770B total parameters with 49B active and pushes context past 1 million tokens. Tencent positions Hy3 as “rivals trillion-scale flagships” because sparse activation lets a ~300B model punch like a much larger dense stack on agent benchmarks without activating every weight [2].
Both models are available on Hugging Face and ModelScope from day one. Hy3’s free API window on OpenRouter (tencent/hy3:free) ran for two weeks at launch; Hy4 preview gets the same two-week free tier on WorkBuddy, CodeBuddy, and OpenRouter [3]. AtlasCloud lists Hy3 at $0.20 per million input tokens and $0.80 per million output tokens with the full 256K context — placing it among the most cost-efficient hosted options [4].
How they stack up against GLM, DeepSeek, and closed APIs
Tencent’s internal blind evaluation had 163 experts score Hy4 preview at 2.99/4.00 across 203 engineering tasks, edging out GLM-5.3 (2.92) and Kimi K3 (2.94) [3]. Hy3 posted 78% on SWE-bench Verified and 90.4% on GPQA Diamond [2]. The Hy3 preview base model already competed with Kimi-K2 Base (1,043B), DeepSeek-V3 Base (671B), and GLM-4.5 Base (355B) while activating only 21B parameters against their 32B–37B [4].
The practical difference: Apache 2.0 at this scale means enterprises that cannot touch US export-controlled APIs get a permissive-license alternative for non-regulated coding agents [2]. Nous Research framed it as “focused on cost-effective agentic use, and particularly strong on coding, tool-calling reliability, reasoning, and 256K long-context tasks” [2].
Running them yourself: API, self-hosting, and quantization
If you want to self-host Hy3 at full fidelity, Tencent specifies 8× GPU tensor parallel using H20-3e or H200-class GPUs [2]. That’s a cluster, not a workstation. However, 1-bit and 4-bit GGUF quantizations for llama.cpp with MTP support arrived July 14, enabling inference on 128 GB single-GPU rigs — no 8× H20 cluster required [2]. Official serving recipes exist for vLLM and SGLang [2].
Hy4 preview’s 49B active parameters and 1M+ context will raise the bar further; Tencent hasn’t published minimum self-host specs for it yet. For now, the free API windows and third-party hosts (OpenRouter, Tencent Cloud TokenHub, AtlasCloud) are the lowest-friction entry points [3][4].
Where they’re already showing up in real workflows
Hy3 powers Tencent’s WorkBuddy/CodeBuddy coding assistants, the Yuanbao consumer assistant (which gained a free Agent function generating PowerPoint, Word, Excel, PDF, and HTML from natural language), the OS-level Marvis assistant, the knowledge tool ima, Weixin/WeChat Official Accounts AI customer service, and the Path of Exile: Advent assistant on WeGame [4]. WorkBuddy users actively selecting Hy3 grew sixfold since the preview [4].
Hy4 preview extends this: stronger understanding, planning, debugging, and validation for long-context development; enhanced financial analysis and cross-document collaboration; playable game prototypes from single natural-language requests; and notable gains in AI R&D, molecular dynamics, condensed-matter physics, and fundamental mathematics [3]. It also participated in its own development — proposing training optimizations, running experiments, and iterating based on results, establishing an early recursive self-improvement loop [3]. Inference throughput improved 31.8% from autonomous operator fusion and communication optimization [3].
What to weigh before switching
The Apache 2.0 license is commercially friendly [5]. The benchmark numbers are competitive. But two practical constraints remain: full-weight self-hosting demands serious GPU capital, and the models are weeks-to-months old — long-term stability, community tooling, and security audit trails are still forming. If your workload fits the free API tier or hosted pricing, the cost argument is strong. If you need air-gapped deployment at full fidelity, budget for the 8× H200-class cluster. The 128 GB single-GPU quantized path exists for Hy3; Hy4’s equivalent isn’t documented yet.
Start with the free API on OpenRouter or WorkBuddy/CodeBuddy. Stress-test your actual agent workflows — tool-calling reliability, long-context retrieval, output-format adherence — before committing infrastructure.
Sources
- Tencent Hy3: 295B Open MoE for Agentic Coding | explainx.ai Blog
- Tencent Releases and Open-Sources Tencent Hy4 preview
- Tencent releases Hy3: open-source 295B MoE model - datanorth.ai
- Tencent Hunyuan Officially Releases Hy3, Advancing Agent Capabilities …
- China’s Tencent releases new open-source AI model for coding, research …
