The first entry is always the pin's source. Overall confidence is a weighted average of how firmly each reference supports the start and end times used above; a reference counts half as much for every 180 days older than the newest.
Tencent's[6] release post is dated 2026.07.06 and headlined "Tencent Hunyuan Officially Releases Hy3"; TechNode's[4][5] report of 7 July says the launch was "On Monday", which is 6 July.
Tencent[1][6] Hy Team's model card: 295B total and 21B active parameters (3.8B in the MTP layer), 256K context, Apache 2.0, BF16 and FP8, the 270-expert blind evaluation (2.67 of 4 against GLM-5.1's 2.51) and the hallucination, commonsense and multi-turn error-rate drops; the benchmark chart against GLM5.2, Seed2.1 Pro, DeepSeek V4 pro, Qwen3.7 Max, GPT 5.5 and Claude Opus 4.8 is its benchmark.png.
Independent index: Hy3 scores 25 on the Artificial Analysis Intelligence Index v4.3.2 against 23 (reasoning) and 17 (non-reasoning) for Hy3-preview; 256k context, about 87 output tokens per second, $0.07 per Intelligence Index task, weights on Hugging Face[2].
What came next: Tencent[1][6] open-sourced the Hy4 preview on 28 August with 770B total and 49B active parameters and a context window of more than 1 million tokens, priced at $0.834 per 1 million input tokens and $2.501 output.
TechNode's[4] 7 July report: the launch was "On Monday"; API at RMB 1 per 1 million input tokens (about $0.15), RMB 4 output (about $0.59) and RMB 0.25 cached input, on Tencent[1][6] Cloud's TokenHub; Hy3 integrated into WorkBuddy/CodeBuddy, Yuanbao, Marvis and ima.
Something missing or wrong? Say it in your own words: a link that backs this pin up, a different start or end date and why, or a fact it lacks or gets wrong. The AI checks it against this pin's sources, searches for better ones, and adds any page that backs you up. The pin's own sources still count most. A picture that shows something else, or shows it badly, is looked at too, and moved down or replaced.
| Benchmark | Hy3 | Hy3 preview | GLM5.2 | Seed2.1 Pro | DeepSeek V4 pro | Qwen3.7 Max | GPT 5.5 | Claude Opus 4.8 |
|---|---|---|---|---|---|---|---|---|
| SWE-bench Pro | 57.9 | 46.0 | 62.1 | 57.5 | 55.4 | 60.6 | — | 69.2 |
| SWE-bench Multilingual | 75.8 | 68.3 | 83.0 | 75.1 | 76.2 | 78.3 | — | 84.4 |
| NL2repo | 45.6 | 35.3 | 48.9 | 47.0 | 41.5 | 47.2 | — | 69.7 |
| Terminal Bench 2.1 | 71.7 | 58.0 | 81.0 | 71.0 | 64.0 | 75.0 | — | 85.0 |
| BrowseComp | 84.2 | 67.1 | 79.3 | 86.2 | 83.4 | — | 84.4 | 84.3 |
| MCP Atlas (public) | 79.1 | 66.1 | 82.6 | 80.6 | 79.7 | 79.6 | — | 84.1 |
| ClawEval (pass^3) | 68.5 | 55.0 | 62.4 | 62.1 | 62.1 | 65.2 | — | 72.1 |
| SkillsBench (text-only) | 55.3 | 29.1 | 51.9 | 53.2 | 40.5 | 46.8 | — | 64.6 |
| HLE (with tools, text-only) | 53.2 | 35.4 | 54.7 | 55.7 | 48.2 | 53.5 | — | 57.9 |
| FrontierScience-Olympiad | 74.8 | 70.0 | 72.5 | 70.1 | 70.0 | 74.3 | 73.8 | — |
| MathArena Apex | 38.7 | 12.6 | 16.8 | 31.3 | 38.3 | 44.5 | 85.4 | — |
| AA-LCR | 73.4 | 66.3 | 73.4 | 70.0 | 71.3 | 70.2 | 76.4 | — |