Tencent model product
Hy3
8B MTP layer, developed by the Tencent Hy Team.
Updated Aug 12, 2026. Default version: Hy3
LLMBoard score71.0Hy3
Coverage80%24 benchmark families
Context window256KTokens
Official input priceN/AOfficial price unavailable
Hy3 Specifications
Technical details for the model's default version.
- Version
- Hy3
- Released
- Jul 6, 2026
- Knowledge cutoff
- Unknown
- Parameters
- 295B
- Context window
- 256K
- Max output
- 64K
- Inputs
- text
- Outputs
- text
- Open weights
- Yes
- License
- Apache 2.0
Hy3 Capability Profile
This profile uses the model's current scored version. Arena ratings and prices are shown separately.
Hy3 Benchmark Results
Benchmark scores for Hy3.
| CL-bench | 0.2 | 1 | 2 | 100.0% | C | |
| CL-bench (Life) | 0.2 | 1 | 1 | 100.0% | C | |
| CMT-Benchmark | 0.4 | 1 | 1 | 100.0% | C | |
| HorizonMath | 0.1 | 1 | 3 | 100.0% | C | |
| Humanity's Last Exam (no tools, text-only) | 0.5 | 1 | 1 | 100.0% | C | |
| PHYBench | 0.8 | 1 | 1 | 100.0% | C | |
| AA-LCR | 0.7 | 2 | 16 | 93.3% | C | |
| ArXivMath | 0.5 | 2 | 2 | 0.0% | C | |
| Humanity's Last Exam (with tools, text-only) | 0.5 | 2 | 2 | 0.0% | C | |
| FrontierScience Olympiad | 0.7 | 3 | 3 | 0.0% | C | |
| IMO-AnswerBench | 0.9 | 3 | 19 | 88.9% | C | |
| SkillsBench | 0.6 | 3 | 8 | 71.4% | C | |
| SuperChem | 0.5 | 3 | 3 | 0.0% | C | |
| USAMO 2026 | 30.2 | 3 | 3 | 0.0% | C | |
| WideSearch | 0.8 | 3 | 9 | 75.0% | C | |
| WildClawBench | 0.5 | 3 | 5 | 50.0% | C | |
| Claw-Eval | 0.7 | 4 | 13 | 75.0% | C | |
| DeepSearchQA | 0.9 | 4 | 9 | 62.5% | C | |
| FrontierScience Research | 0.2 | 4 | 4 | 0.0% | C | |
| MathArena Apex | 0.4 | 4 | 7 | 50.0% | C | |
| NL2Repo | 0.5 | 6 | 14 | 61.5% | C | |
| APEX-Agents | 0.3 | 7 | 7 | 0.0% | C | |
| MCP Atlas | 0.8 | 7 | 31 | 80.0% | C | |
| DeepSWE | 0.3 | 9 | 10 | 11.1% | C | |
| SWE-bench Multilingual | 0.8 | 9 | 34 | 75.8% | C | |
| Terminal-Bench 2.1 | 0.7 | 13 | 19 | 33.3% | C | |
| BrowseComp | 0.8 | 14 | 58 | 77.2% | C | |
| Toolathlon | 0.5 | 17 | 31 | 46.7% | C | |
| SWE-Bench Pro | 0.6 | 19 | 45 | 59.1% | C | |
| SWE-Bench Verified | 0.8 | 20 | 105 | 81.7% | C | |
| GPQA | 0.9 | 21 | 234 | 91.4% | C |
Hy3 Arena Results
Preference and agent-evaluation results for the default version.
| agent praise complaint | overall | 19 | 0.0 | N/A | 2K | |
| webdev | overall | 20 | 1524.2 | 2,175 | N/A | |
| agent bash recovery steps | overall | 28 | 0.0 | N/A | 12.6K | |
| agent | overall | 30 | -0.0 | N/A | 477.6K | |
| agent task outcome explicit | overall | 31 | -0.0 | N/A | 6.7K | |
| agent steerability | overall | 39 | -0.1 | N/A | 8.1K | |
| agent tool hallucination | overall | 41 | -0.0 | N/A | 448.2K | |
| text style control | overall | 52 | 1456.3 | 4,544 | N/A | |
| text | overall | 58 | 1441.0 | 4,544 | N/A | |
| text factuality | overall | 59 | 1448.0 | 4,544 | N/A | |
| webdev | overall | 79 | 1356.3 | 1,394 | N/A | |
| text factuality | overall | 117 | 1415.7 | 6,624 | N/A | |
| text style control | overall | 122 | 1412.8 | 6,627 | N/A | |
| text | overall | 127 | 1405.0 | 6,627 | N/A |
Hy3 Pricing
Official vendor API pricing appears first, followed by individual provider offers.
- Official API
- N/A
- Official provider
- N/A
- Lowest third-party
- From $0.066 input, $0.26 output per 1M via NanoGPT
- Tracked offerings
- 11
| Tencent TokenHub | hy3 | global | N/A | N/A | 256K | |
| Tencent Token Plan | hy3 | global | N/A | N/A | 256K | |
| NanoGPT | tencent/hy3 | global | $0.066 | $0.26 | 262.1K | |
| OpenRouter | tencent/hy3 | global | $0.132 | $0.528 | 262.1K | |
| OpenCode Go | hy3 | global | $0.14 | $0.58 | 256K | |
| LLM Gateway | hy3 | global | $0.14 | $0.58 | 262.1K | |
| Kilo Gateway | tencent/hy3 | global | $0.14 | $0.58 | 262.1K | |
| Vercel AI Gateway | tencent/hy3 | global | $0.14 | $0.58 | 262.1K | |
| Hugging Face | tencent/Hy3 | global | $0.14 | $0.58 | 262.1K | |
| Deep Infra | tencent/Hy3 | global | $0.14 | $0.58 | 262.1K | |
| CrossModel | tencent/hy3 | global | $0.16 | $0.64 | 262.1K |
Official prices use only the vendor's configured official Provider and positive standard USD PAYG rates. Third-party offers remain explicitly labeled.
Hy3 Versions
Available versions of this model. The score column identifies the version used in the overall ranking.
| Hy3 | 71.0 | 295B | 256K | 64K | Yes | Apache 2.0 |
What is Hy3?
Key information about Hy3 and its available data.
8B MTP layer, developed by the Tencent Hy Team. Following the Hy3 Preview launch in late April, the team scaled up post-training with higher-quality data and RL, gathering feedback from 50+ products.
Hy3 outperforms similar-size models and rivals flagship open-source models with 2-5x the parameters, with strong gains in reasoning, agentic, and long-context tasks.
It uses 80 layers (plus 1 MTP layer), 64 GQA attention heads (8 KV heads, head dim 128), a 4096 hidden size, 192 experts with top-8 activated, a 256K context window, and BF16 precision.
Hy3 is a hybrid-thinking model supporting configurable reasoning effort (no_think, low, high), and emphasizes production-grade tool-call and output-format stability, reduced hallucination, and reliable multi-turn intent tracking.
Data as of 2026-08-11.
Hy3 vs nearby models
Open a comparison with the three ranked models immediately above and below this model.
Models similar to Hy3
Recommendations prioritize the same model type and family, then the closest LLMBoard score.
MI
MiniMax M3
MiniMax
OP
GPT-5.2
OpenAI
OP
GPT-5.2-Pro
OpenAI
ZA
GLM 5.1
Zhipu AI
BY
Seed 2.0 Pro
ByteDance
AN
Claude Opus 4.5
Anthropic
FAQ
Common questions about Hy3.
When was Hy3 released?
Hy3's default version was released on Jul 6, 2026.
How much does Hy3 cost?
No official standard PAYG price is currently available for Hy3. The lowest tracked third-party offer starts at $0.066 input and $0.26 output via NanoGPT.
Who created Hy3?
Hy3 was created by Tencent.
What is the context window for Hy3?
The default version has a 256K token context window.
Is Hy3 open weight?
Yes. The default version is marked as open weight under Apache 2.0.
How many API providers offer Hy3?
11 provider offerings are linked to the default version.
What models should I compare Hy3 with?
Nearby ranked alternatives include GPT-5.2-Pro, GPT-5.2, MiniMax M3.