Tencent Releases Hy3: 295B Open MoE Model Under Apache 2.0
Tencent has released Hy3, the full production version of its Hunyuan reasoning-and-agent model, following up on the Hy3 Preview that launched in late April 2026. Hy3 is a 295-billion-parameter Mixture-of-Experts (MoE) model with 21 billion active parameters, and it ships fully open-source under the Apache 2.0 license — weights, quantized variants, and inference recipes included.
Intermediate
What Changed Since the Preview
Tencent says the Hy Team collected feedback from more than 50 products running Hy3 Preview — including its CodeBuddy and WorkBuddy agent tools — and used it to scale up post-training with higher-quality data. The result is a model the company describes as “reliable and cost-effective” rather than simply bigger: total parameters actually shrank from earlier 400B+ Hunyuan models to what Tencent calls a sweet spot between capability and inference cost.
Reliability metrics moved the most between preview and release. Tencent reports the hallucination rate fell from 12.5% to 5.4%, and the multi-turn conversation issue rate dropped from 17.4% to 7.9%, both attributed to fine-grained data cleaning and tighter training constraints. In blind evaluations by 270 outside domain experts across real-world tasks, Hy3 scored 2.67 out of 4, ahead of GLM-5.1’s 2.51, with particular strength in frontend development and CI/CD workflows.
Architecture and Specs
Hy3 uses a 192-expert MoE design with the top 8 experts activated per token, spread across 80 transformer layers plus a single multi-token-prediction (MTP) layer for speculative decoding. Other specs: 64 attention heads using grouped-query attention (8 KV heads, 128 head dimension), a 4096 hidden size, 13,312 intermediate size, a 120,832-token vocabulary, and a 256K-token context window. The model is distributed in standard BF16 and an FP8-quantized variant, with day-one support for vLLM and SGLang.
Tencent’s own benchmark chart shows Hy3 improving substantially over Hy3 Preview across the board — for example, SWE-bench Pro climbs from 46.0 to 57.9, Terminal Bench 2.1 from 58.0 to 71.7, and BrowseComp from 67.1 to 84.2. Against the current field of frontier models, Hy3 lands roughly mid-pack: it trails Claude Opus 4.8 and GPT 5.5 on most reasoning and coding benchmarks, but is competitive with or ahead of GLM5.2, Seed2.1 Pro, DeepSeek V4 pro, and Qwen3.7 Max on several, including BrowseComp (84.2) and AA-LCR (73.4). On MathArena Apex, Hy3 still trails the pack (38.7) behind GPT 5.5’s 85.4, underscoring that its strengths lean toward agentic coding and long-context tasks rather than competition math.
Cost and Availability
Tencent is pricing API access aggressively: roughly $0.18 per million input tokens and $0.59 per million output tokens, alongside a 40% inference-efficiency improvement and reported 54% reduction in time-to-first-token for its internal CodeBuddy and WorkBuddy products. Hy3 has been validated on stable agent runs of up to 495 steps in production traffic. Weights are available on Hugging Face, ModelScope, GitCode, and CNB, and the model is also listed on OpenRouter and Tencent Cloud’s TokenHub.
What This Means
Hy3 is another entry in a fast-moving lineup of large, permissively-licensed Chinese MoE models — following Meituan’s LongCat-2.0 (1.6T/48B active, MIT), Xiaomi’s MiMo-V2.5-Pro (1T/42B active), and DeepSeek’s V4 — but it takes a different approach by prioritizing production reliability and cost per step over raw parameter count. At 295B total and 21B active, Hy3 is comparatively lean next to its trillion-parameter peers, betting that consistent agent behavior across hundreds of tool-call steps matters more for real deployments than topping a leaderboard.
Related Coverage
- Tencent Open-Sources HY-World 2.0: A Multi-Modal 3D World Model — Tencent’s most recent prior open-source release, a 3D world model framework.
- Meituan Open-Sources LongCat-2.0, a 1.6T Model Trained on Chinese Chips — a competing large-scale Chinese MoE release from the same period.
- DeepSeek Releases V4: Open-Source 1.6T MoE with 1M Context — another major open-weight MoE model Hy3 is benchmarked against.
沪公网安备31011502017015号