Tencent Hy3 Preview, open-sourced April 23, 2026, is a 295B-parameter Mixture-of-Experts model with 21B active parameters, fusing fast-and-slow thinking in a single model. It scores 74.4% on SWE-bench Verified and 54.4% on Terminal-Bench 2.0 (self-reported), with 256K context and demonstrated production deployments of 495-step agentic workflows in Tencent's own CodeBuddy and WorkBuddy products.
Why It's in Assess
Hy3 Preview represents Tencent's first major open-weight release after rebuilding its pre-training and reinforcement-learning infrastructure from scratch in early 2026 — reaching open-source release in under three months from cold start. That pace is impressive, but the benchmarks are self-reported and the model family is untested beyond Tencent's own internal products:
- 74.4% SWE-bench Verified (self-reported), closing ground on Kimi K2.6 (80.2%) and GPT-5.5 (82.6%) but without independent reproduction
- Strong agentic production evidence: In CodeBuddy and WorkBuddy deployments, Hy3 Preview achieved 99.99% success rate over 495-step agent workflows, a 54% reduction in time-to-first-token, and 47% lower end-to-end response time — but these figures are internal Tencent measurements
- Cost-efficient architecture: 21B active parameters from a 295B total pool makes inference significantly cheaper than dense models of similar reported capability
- Rebuilt infrastructure: Tencent reset its AI stack entirely in early 2026, meaning Hy3 Preview reflects the first output of the new pipeline — early days, but with serious engineering investment behind it
Move to Trial when: independent evaluators (Princeton SWE-bench.com, Artificial Analysis, or equivalent) confirm the SWE-bench and Terminal-Bench scores.
Capabilities
- Context: 256K tokens
- Reasoning modes: Fast-and-slow thinking fused into a single model (no separate chain-of-thought variant required)
- Agentic tool use: MCP toolchain orchestration; document processing, data analysis, knowledge retrieval
- Long-horizon task completion: 495-step autonomous workflows validated in production
- Benchmark performance (all self-reported):
| Benchmark | Hy3 Preview | Notes |
|---|---|---|
| SWE-bench Verified | 74.4% | Self-reported |
| Terminal-Bench 2.0 | 54.4% | Self-reported |
| BrowseComp | 67.1% | Self-reported |
| WideSearch | 70.2% | Self-reported |
Cautions
- All public benchmarks are self-reported by Tencent; no independent reproductions as of May 2026
- Same data sovereignty considerations as other Chinese-origin models (Tencent is subject to Chinese data law)
- "Preview" status signals the family is not yet at production release quality for general use
- Limited Western ecosystem presence and community adoption compared to DeepSeek or Kimi
Key Characteristics
| Property | Value |
|---|---|
| Provider | Tencent |
| License | Hy3 Preview License |
| Pricing | Open weights (free to download and run) |
| Context window | 256,000 tokens |
| Parameters | 295B total / 21B active (MoE) |
| Architecture | Hybrid fast-and-slow MoE; 3.8B MTP layer parameters |
| Status | Preview (April 23, 2026) |
| GitHub | Tencent-Hunyuan/Hy3-preview |
| Website | hy3ai.com |
Further Reading
- Tencent Unveils Hy3 Preview announcement (Tencent)
- Tencent Open-Sources Hy3 at 74.4% SWE-bench (Business Analytics Substack)
- Tencent Uses Product Rollout, Not Just Benchmarks, To Define Hy3 Preview (Dataconomy)
- Tencent's New Hy3 AI Model Is the Most Efficient Chinese LLM No One's Talking About (Decrypt)
- Hy3-preview on Hugging Face