Tencent's Hy3 Beats GLM-5.2 Everywhere But Code — and Apache 2.0 Changes the Game

Tencent has released Hy3, a 295B-parameter MoE model, under the permissive Apache 2.0 license — removing the legal blockers that kept many Chinese open-weight models out of Western enterprise deployments. It outperforms GLM-5.2 on agentic search and tool orchestration, though trails on coding benchmarks. The deployment economics and reliability improvements may matter more to production teams than leaderboard rankings.

Tencent's Hunyuan team has officially released Hy3, a 295-billion-parameter Mixture-of-Experts model, under the Apache 2.0 license — a deliberate reversal from its April preview release and a significant shift for enterprise adoption. The move removes the regional license restrictions that had blocked deployments serving the EU, UK, and South Korea, the fine print that legal teams have routinely used to kill open-weight model evaluations before they start.

The open-model community on X reacted immediately, with researchers flagging the license change as the actual headline — one widely circulated post argued that if the benchmark scores hold up, Tencent has become one of the leaders of open source. Tencent is also offering free access on OpenRouter for two weeks.

From Preview to Production in Ten Weeks

Hy3's April preview was the first model off Tencent's rebuilt pre-training and reinforcement learning infrastructure, shipped fewer than three months after a February overhaul. Chief AI Scientist Shunyu Yao framed the early release as a feedback-gathering exercise — and Tencent says that's exactly how it played out.

The team collected input from more than 50 internal product teams, fixed issues in task execution and multi-turn interaction, and scaled up its post-training pipeline. The architecture itself is unchanged:

295B total parameters, 21B active per forward pass
Top-8 routing across 192 experts
3.8B-parameter multi-token prediction (MTP) layer for speculative decoding
256K context window

What changed is behavior — and, critically, reliability.

Benchmark Wins and One Clear Loss

Tencent's headline evaluation is a blind human study rather than a standard leaderboard submission. 270 domain experts worked on real-world workflows, generating 312 valid comparisons. Hy3 scored 2.67 out of 4 against GLM-5.1's 2.51, with the strongest advantages in frontend development, CI/CD, and data infrastructure tasks.

The choice of opponent matters here. Zhipu AI released GLM-5.2 in mid-June, and Tencent's own benchmark appendix shows GLM-5.2 leading on every major agentic coding metric:

| Benchmark | GLM-5.2 | Hy3 | |---|---|---| | SWE-bench Verified | 84.2 | 78.0 | | SWE-bench Multilingual | 83.0 | 75.8 | | Terminal-Bench 2.1 | 81.0 | 71.7 | | DeepSWE | 46.2 | 28.0 |

The gap is less surprising when you size the models side by side. GLM-5.2 is a ~744B-parameter MoE with ~40B active parameters per token; Hy3 runs at fewer than half the total parameters and roughly half the per-token compute. Tencent is trailing on coding with a significantly smaller model.

Hy3's genuine wins sit in search-heavy and tool-orchestration workloads:

BrowseComp: 84.2 — ahead of all open models in Tencent's table, competitive with Claude Opus 4.8 and GPT-5.5
DeepSearchQA: 91.0
MCP-Atlas (tool orchestration): 79.1 — leads the open field
AA-LCR (long-context retrieval): 73.4

One caveat: nearly all competitor scores in Tencent's appendix are marked as Tencent's own test runs. Independent verification from indices like Artificial Analysis was still pending at publication.

The Reliability Story Enterprises Actually Care About

The more revealing section of the Hy3 model card isn't the benchmark table — it reads more like a production reliability report.

In internal evaluations on real-world scenarios, Tencent reports:

Hallucination rate cut from 12.5% → 5.4%
Commonsense error rate cut from 25.4% → 12.7%
Multi-turn issue rate cut from 17.4% → 7.9%
MRCR long-dialogue benchmark score improved from 42.9% → 75.1%

Tencent attributes these gains to fine-grained data cleaning and training constraints built around an explicit behavior pattern: answer when grounded, disclose when evidence is missing, don't conflate sources, don't fabricate.

Tencent also emphasizes consistency across agent frameworks — reporting SWE-bench variance within a few points whether the model runs inside Claude Code-style harnesses, Cline, or KiloCode. That's an underrated production property: enterprises rarely control which agent scaffold their teams standardize on, and a model that only performs in one harness is a hidden integration cost.

These are self-reported internal measurements and deserve appropriate skepticism. But the choice to foreground them at all signals Tencent's intended customer: teams that have been burned by models that demo well and fabricate confidently in production.

The Deployment Math: Half the Size, Constraint-Optimized Silicon

The economics of self-hosting make Hy3's coding gap against GLM-5.2 look less like a loss and more like a deliberate architectural trade.

GLM-5.2 in FP8 occupies roughly 744GB of GPU memory — making an 8×H200 node the practical minimum for production serving. Hy3's FP8 footprint comes in under 300GB, leaving room for KV cache and batching on far more attainable hardware.

Tencent's recommended serving configuration targets Nvidia's H20-3e — the memory-boosted variant of the H20, the chip Nvidia designed specifically to comply with U.S. export restrictions on China. There is no mention of Huawei or Ascend chips in the deployment guide.

The implication: Hy3 is sized so that eight of the chips Chinese companies can legally buy serve it comfortably at full precision. That constraint-driven design has a convenient side effect for everyone else — a model tuned to run well on deliberately capped silicon runs even more comfortably on H100s, H200s, and B200s available in Western data centers, through standard vLLM and SGLang deployments with MTP speculative decoding.

What This Actually Means

Apache 2.0, no regional exclusions, under-300GB FP8 footprint, and benchmark-verified strength in agentic search and tool orchestration — Hy3 is positioned not as the best open-weight model in absolute terms, but as the most deployable one for a specific production profile.

For teams running search-heavy agents or tool-orchestration pipelines on hardware they control, the case is clear. For teams whose primary workload is repository-scale coding, GLM-5.2 still owns that benchmark — and the gap is wide enough that it isn't likely to close at Hy3's current scale.