Tencent Hunyuan Hy3: A 295B Open-Source Agentic AI for Developers

Tencent Ships Hunyuan Hy3 on 6 July 2026

On 6 July 2026, Tencent officially released Hunyuan Hy3, a 295-billion-parameter open-weight Mixture-of-Experts model under the Apache 2.0 licence. The release completed a process that began with a restricted April 2026 preview under a bespoke Tencent Hy Community Licence Agreement, which explicitly excluded the EU, UK, and South Korea from its usage terms; the July full release removed all geographic restrictions. Hy3 is available on Hugging Face and ModelScope from launch day, as well as via the Tencent Cloud TokenHub API. It is priced at RMB 1 — approximately $0.15 — per million input tokens and RMB 4 — approximately $0.59 — per million output tokens, placing it among the most affordable frontier-adjacent models available to developers globally.

Architecture: 295B Total, 21B Active, Hybrid Thinking

Hy3 uses a sparse Mixture-of-Experts architecture with 295 billion total parameters and 21 billion active parameters per token. It also includes a multi-token-prediction layer carrying an additional 3.8 billion parameters. The model supports a 256,000-token context window. Its most distinctive architectural feature is a hybrid fast-and-slow-thinking design: the model can switch between a rapid-response mode for lower-latency tasks and an extended chain-of-thought reasoning mode for complex agentic operations. This dual capability allows developers to select the inference cost and latency profile appropriate to their use case — fast mode for high-volume interactive tasks, slow reasoning mode for long-horizon autonomous workflows where quality per task matters more than throughput.

The Multi-Token-Prediction Layer

The MTP layer allows Hy3 to predict several tokens simultaneously rather than purely sequentially, increasing generation throughput without a proportionate increase in memory overhead at inference time. For developers running Hy3 on constrained hardware or targeting interactive latency targets, the MTP layer provides speed benefits that standard autoregressive decoding alone would not deliver at 21 billion active parameters.

Benchmarks: Matching Models Two to Five Times Its Active Size

Tencent's benchmark results show Hy3 performing near or above models carrying two to five times its number of active parameters. On Tencent's WorkBuddy enterprise platform, Hy3 achieved a 90 per cent task resolution rate across agentic workflows spanning web research, data extraction, code generation, and document processing — a practical production metric rather than a held-out test-set score. In scientific reasoning evaluations, Hy3 matches or exceeds GPT-5.5 while carrying a fraction of the per-token inference cost. Comparisons with Claude Opus 4.8 and DeepSeek V4 show Hy3 as competitive across a range of tasks while substantially cheaper in both self-hosted and cloud-API configurations.

Pricing and Availability

Hy3 is accessible through three routes. The full open weights are freely downloadable from Hugging Face and ModelScope under Apache 2.0, with no usage fees and no geographic restrictions. The Tencent Cloud TokenHub API provides managed inference at RMB 1 per million input tokens, RMB 4 per million output tokens, and RMB 0.25 per million cached input tokens. Tencent has also committed to progressive integration with global third-party developer platforms. At the API pricing — $0.15 per million input tokens and $0.59 per million output tokens — Hy3 costs a fraction of what frontier-proprietary models from OpenAI and Anthropic charge at comparable capability levels.

The Developer Case: Open Weights at Low Cost

For engineering teams integrating AI into production applications, Hy3's combination of open weights, Apache 2.0 licensing, 256K context, and low API pricing addresses three common barriers simultaneously. Open weights and permissive licensing allow self-hosted deployment, eliminating per-token API costs and data residency concerns. The 256K context window accommodates large-codebase analysis, long-document workflows, and multi-step agentic task chains within a single model call — reducing the need for chunking pipelines and retrieval-augmented generation architectures that add latency and engineering complexity. The 90 per cent agentic task resolution rate on WorkBuddy speaks directly to production reliability, which is what teams integrating AI into enterprise software actually need to evaluate.

What Hy3 Means for Indian Engineering Teams

Indian software product companies and IT services providers building AI features for enterprise clients face a set of constraints that Hy3 addresses directly. The first is cost at scale. At $0.15 per million input tokens, a team processing 10 billion input tokens per month — a realistic volume for a mid-size AI product — pays approximately $1,500 in input costs, compared to $30,000 or more on models like GPT-5.6 Sol. For teams that self-host the open weights on Indian data centre infrastructure, the per-token API cost disappears entirely.

The second is data localisation. Indian banking, healthcare, and government-adjacent applications face tightening data residency requirements under the Digital Personal Data Protection Act and sector-specific RBI and SEBI frameworks. Hy3's open weights allow these teams to deploy on-premise or in Indian cloud regions, keeping all data within jurisdiction.

The third is the context window. Processing a full insurance policy, a multi-year loan repayment history, or a large enterprise codebase in a single 256K-token context avoids the quality degradation and latency overhead of retrieval-augmented chunking — a practical advantage for teams building document-intensive or code-intensive agentic products in Indian fintech, legal-tech, and enterprise SaaS.

The Bottom Line

Tencent released Hunyuan Hy3 on 6 July 2026: a 295-billion-parameter open-weight Mixture-of-Experts model with 21 billion active parameters, a 3.8-billion-parameter MTP layer, and a 256K context window. It is available under Apache 2.0 on Hugging Face and ModelScope with no geographic restrictions, and via Tencent Cloud TokenHub at $0.15 per million input tokens and $0.59 per million output tokens. It achieves a 90 per cent agentic task resolution rate on Tencent's enterprise platform and benchmarks near or above models carrying two to five times its active parameter count. For engineering teams building AI-powered products, Hy3 offers frontier-adjacent capability at open-source pricing — and the self-hosting option gives teams in regulated sectors a route to capable agentic AI without routing data through overseas API providers.