For the previous 12 months, the awkward secret of the open-weight mannequin growth has been that most of the strongest Chinese language releases have been off-limits to a big slice of the enterprises most enthusiastic about them. License phrases that excluded the European Union, the UK and South Korea meant authorized groups killed deployments earlier than engineering groups completed their evals — not only for corporations headquartered there, however for any enterprise serving visitors into these areas. For IT groups weighing open fashions, the trade-offs are unusually express.
Tencent simply eliminated that impediment. The corporate's Hunyuan workforce launched the total model of Hy3, a 295-billion-parameter Combination-of-Consultants (MoE) mannequin with 21 billion lively parameters, and — in a reversal from April's preview launch — shipped it beneath the permissive Apache 2.0 license. The response from the open-model group was quick, with researchers on X singling out the license change as the actual headline, and one extensively shared publish arguing that if the scores maintain up, Tencent has simply change into one of many leaders of open supply. Tencent says it is going to be free on OpenRouter for 2 weeks.
The scores are value scrutinizing — and so they don't all level the identical course. However the extra attention-grabbing story is what Tencent selected to steer with: reliability metrics and deployment economics aimed squarely at manufacturing use.
From preview to product in ten weeks, formed by 50 inside groups
Hy3's April preview was the primary mannequin of Tencent's rebuilt pre-training and reinforcement studying infrastructure, shipped lower than three months after the February rebuild. Chief AI Scientist Shunyu Yao framed the early open launch as a deliberate transfer to assemble suggestions from builders and customers earlier than the official model — and Tencent says that's precisely what occurred. In accordance with the mannequin card, the workforce collected suggestions from greater than 50 product groups after the late-April preview, fastened points in process execution and interplay, and scaled up its post-training pipeline.
The structure is unchanged: 295B whole parameters, 21B lively per ahead cross through top-8 routing throughout 192 specialists, a 3.8B-parameter multi-token prediction (MTP) layer for speculative decoding, and a 256K context window. What modified is habits. Tencent's positioning is that the total launch considerably outperforms similar-size fashions and rivals flagship open-source fashions with two to 5 occasions the parameters.
That "two to 5 occasions" framing is sensible for the place this mannequin is aimed — and it invitations a direct comparability with the present open-weight coding chief, GLM-5.2.
Tencent's blind check favors Hy3 over GLM-5.1, however GLM-5.2 nonetheless owns coding
Tencent's headline analysis is a blind human research quite than a leaderboard. Arguing that public benchmarks don't inform the total story, the corporate ran a blind check with 270 specialists throughout disciplines engaged on real-world workflows, amassing 312 legitimate comparisons, through which Tencent reviews that Hy3 scored 2.67 out of 4 in opposition to GLM-5.1's 2.51 — with the clearest benefits in frontend improvement, CI/CD, and knowledge and storage work.
The selection of opponent issues. Zhipu AI launched GLM-5.2 in mid-June, and Tencent's personal benchmark appendix reveals GLM-5.2 forward of Hy3 throughout primarily the complete agentic coding suite: SWE-bench Verified (84.2 vs. 78.0), SWE-bench Multilingual (83.0 vs. 75.8), Terminal-Bench 2.1 (81 vs. 71.7) and DeepSWE by a large margin (46.2 vs. 28.0). The blind check focused the older mannequin; the newer one retains the coding crown.
GLM-5.2's coding lead is much less shocking when you contemplate the sizes are aspect by aspect: GLM-5.2 is roughly a 744-billion-parameter MoE with round 40 billion lively parameters per token, in opposition to Hy3's 295 billion whole and 21 billion lively. Tencent is fielding a mannequin with lower than half the parameters — and almost half the per-token compute — of the one it trails.
Hy3's real wins sit elsewhere. On agentic search, it posts 84.2 on BrowseComp and 91.0 on DeepSearchQA — forward of each open mannequin in Tencent's desk and aggressive with Claude Opus 4.8 and GPT-5.5. It leads the open subject on instrument orchestration (79.1 on the general public MCP-Atlas set), on agent-harness evaluations like ClawEval, and on long-context retrieval (73.4 on AA-LCR). Learn collectively, the appendix suggests a mannequin that’s arguably the very best open-weight selection for search-and-tool-heavy agent workloads, whereas conceding repository-scale coding to GLM-5.2.
One caveat applies to each the wins and the losses: almost all competitor numbers in Tencent's appendix are marked as coming from Tencent's personal check runs. Impartial verification, from indices like Synthetic Evaluation, continues to be pending as of publication.
The reliability pitch: hallucination charges lower in half
The place the discharge will get most attention-grabbing for enterprise consumers is the set of numbers Tencent selected to emphasise as a substitute of benchmarks. The mannequin card reads much less like a leaderboard announcement and extra like a manufacturing reliability report.
In inside evaluations on real-world eventualities, Tencent says Hy3's hallucination price dropped in comparison with the preview model from 12.5% to five.4%, and commonsense error charges fell from 25.4% to 12.7% — enhancements it attributes to fine-grained knowledge cleansing and coaching constraints constructed round an express habits sample: reply when grounded, state when proof is lacking, don't conflate sources, don't fabricate knowledge. Multi-turn habits will get the identical therapy: the difficulty price on inside multi-turn checks fell from 17.4% to 7.9%, and Tencent reported that the mannequin's rating on the open MRCR long-dialogue benchmark jumped from 42.9% to 75.1%.
Tencent additionally emphasizes consistency throughout agent scaffolds — reporting SWE-bench variance inside a number of factors whether or not the mannequin runs inside Claude Code-style harnesses, Cline or KiloCode. That's an underrated property: enterprises not often management which agent framework their groups standardize on, and a mannequin that solely performs in a single harness is a hidden integration value. These are self-reported inside measurements, and so they deserve the identical skepticism as any vendor benchmark. However the option to foreground them in any respect indicators who Tencent believes its buyer is: groups which have been burned by fashions that demo properly and fabricate confidently in manufacturing.
The deployment math: a 295B mannequin in a 744B world — on export-compliant silicon
The reliability story connects on to the economics, and that is the place Hy3's coding hole in opposition to GLM-5.2 begins to appear like a deliberate commerce quite than a loss.
GLM-5.2 is a roughly 744-billion-parameter MoE with about 40 billion lively parameters per token; in FP8, its weights alone devour roughly 744GB, making an 8x H200 node the sensible minimal for manufacturing serving. Hy3, at 295B whole parameters, carries an FP8 footprint of beneath 300GB — lower than half the reminiscence, with roughly half the lively parameters per token driving decrease per-request compute. For a company deciding what to self-host, that's the distinction between one heavily-specced node and one thing way more attainable, with room left over for KV cache and batching.
There's a geopolitical wrinkle within the deployment information value noticing too: Tencent's advisable serving configuration targets Nvidia’s H20-3e — the memory-boosted variant of the H20, the GPU Nvidia designed particularly to adjust to U.S. export restrictions on China. Not like GLM-5.2, there isn’t a point out of Huawei or Ascend chips right here. In different phrases, the mannequin is sized in order that eight of the chips Chinese language corporations can legally purchase comfortably serve it at full precision. That constraint-driven design has a handy aspect impact for everybody else: a mannequin that runs properly on intentionally capped silicon runs much more comfortably on the H100s, H200s and B200s accessible in Western knowledge facilities, by commonplace vLLM and SGLang deployments with MTP speculative decoding.
Add the Apache 2.0 license — no regional exclusions, no field-of-use restrictions — and the enterprise equation turns into clear. GLM-5.2 stays the open-weight selection when coding efficiency is the one criterion and an 8x H200 finances is out there. Hy3 makes its case all over the place else: search and tool-heavy agent workloads, reliability-sensitive purposes and organizations that need frontier-adjacent functionality with out frontier-scale infrastructure. The open query is whether or not Western enterprises, now that the license barrier is gone, will deal with a Tencent mannequin as a severe candidate in any respect — or whether or not the following Synthetic Evaluation replace settles the benchmark debate earlier than procurement will get the possibility.
