T
Hunyuan 3 Preview
Open SourceTencent•Released on 2026-04-23
Tencent's first open-source release after a full rebuild of its pretraining and RL infrastructure. 295B MoE with 21B active parameters, 256K context, and hybrid fast/slow thinking. Scores 74.4 on SWE-Bench Verified and 87.2 on GPQA Diamond, with notable gains in coding and long-horizon agent tasks (drives up to 495-step tool-use chains). Led by chief AI scientist Shunyu Yao (ex-OpenAI, ReAct and Tree-of-Thoughts contributor).
84
Overall Score
Core Specs
262K
Context Window
33K
Max Output
ReasoningOpen Sourcetext
Scenario Scores
Pros & Cons
Sentiment0% +0% ·0% −
Pros
- +Open weights (Tencent Hy Community License)
- +295B MoE with only 21B active — large-model quality at mid-model inference cost
- +256K native context
- +Strong coding + agent benchmarks (SWE-Bench 74.4, TerminalBench 54.4)
- +Multi-token prediction (MTP) layer for speculative decoding
Cons
- −Released <24h ago — real-world feedback thin
- −Not fully Apache-licensed — community license has commercial-scale restrictions
- −Deployment requires 8× H20-3e-class GPUs for full precision
- −Text-only (no vision)
Pricing
Input (per 1M tokens)$0.00
Output (per 1M tokens)$0.00
Free trial available
Updated on 2026-04-24
Get Started
1Visit the provider's website
2Create an account
3Start using the model
Benchmarks
sweBenchVerified74.4%
terminalBench254.4%
gpqaDiamond87.2%
hle30%
mmlu87.42%
mmluPro65.76%
math76.28%
gsm8k95.37%
liveCodeBenchV634.86%
Reliability
Incidents (30d)0