Hy
by Tencent · Model · Language & code models
mixed
Strong and cheap on security tasks, but one user saw broken responses.
58% on AA Cyber Index (#3) and 76% on CWE-Bench-AA (#1) at $0.92 per task
AccessOpen weightsNeeds~229 GBDownloadabout 229 GB for the full Hy4 preview weightsLicenseApache 2.0
Why we say so
Artificial Analysis measured Hy at 58% on its Cyber Index (#3) and 76% on CWE-Bench-AA (#1), at $0.92 per task. An HN commenter using it via OpenRouter reported many broken responses in a basic error-correction test, which is a single voice. Tencent's own claims for Hy4 and Hy4 Flash are the weakest evidence, and Tencent concedes Flash trails Hy4 and DeepSeek on terminal coding.
Versions
| Version | Released | Rank |
|---|---|---|
| Hy4LATEST unproven Only Tencent's own claims so far, and the weights are a preview | Aug 28 | |
| Hy3 | Jul 16 |
Evidence
- Nov 17Via OpenRouter, produced many broken responses in one user's basic error-correction test
- Nov 8Tencent Hy4 adds visual understanding and computer use; strong coding, 3D and knowledge work; usable via Tencent Cloud and WorkBuddy
- Aug 30Tencent opens Hy4 Preview weights: 770B MoE, 49B active, gated DeepSeek sparse attention, Apache 2.0, 1.56TB
- Aug 30Free through the API for the first 2 weeks; heavy load caused heavy throttling, so the test took over an hour
- Aug 30Beats Qwen 3.8 Max and DeepSeek V4 Pro, and is close to GLM 5.3 and Kimi K3
Show all 13
- Aug 30Ties GLM 5.3 Flash on Arena web dev leaderboard
- Aug 28In Tencent's blind evaluation with 163 experts, Hy4 preview edged out GLM 5.3 and Kimi K3
- Aug 28Claimed to lead Terminal-Bench 2.1 and DeepSWE-style agentic tests, ahead of GLM 5.3, Kimi K3 and DeepSeek V4 Pro
- Aug 28Claimed to trade the lead with Claude Opus 5 depending on the task
- Aug 28Given an unhinted broken Docker app (Flask, Postgres, frontend) via Hermes Agent, it found the root bug and applied the right fix; results verified in the UI
- Aug 28Architecture adds Identity Hyper Connections (parallel residual lanes) and Gated DSA (selective sparse attention reused across layers) for long context
- Jul 22Tencent released Hy3 295B MoE in 1-bit GGUF at about 83 GB, served on a single GPU
- Jul 22Per token it routes 8 of 192 experts per layer (15,168 experts total); about 5 GB of dense attention weights stay in RAM