Tencent's Hy3 challenges larger AI rivals at half the parameter count

Tencent's new Hy3 model matches or beats flagship open-weight AI models twice its size on agent and search tasks, while cutting deployment costs by more than half.

Tencent's Hunyuan Hy3, a 295-billion-parameter Mixture-of-Experts model with just 21 billion active parameters, matches or beats flagship open-weight models twice its size on agent and search benchmarks — while carrying an Apache 2.0 license and an FP8 memory footprint under 300 GB.

"The model is sized so that eight of the chips Chinese companies can legally buy comfortably serve it at full precision," the Hunyuan team said in the model card, referencing Nvidia's H20-3e — the GPU designed to comply with US export restrictions on China.

Hy3's 192-expert architecture activates only 21 billion of its 295 billion total parameters per forward pass, compared with Zhipu AI's GLM-5.2 at roughly 744 billion total and 40 billion active. On agentic search, Hy3 posts 84.2 on BrowseComp and 91.0 on DeepSearchQA — ahead of every open model in Tencent's internal table and competitive with Claude Opus 4.8 and GPT-5.5. It leads the open field on tool orchestration with 79.1 on MCP-Atlas and on long-context retrieval with 73.4 on AA-LCR. The trade-off is coding: GLM-5.2 still leads on SWE-bench Verified (84.2 vs 78.0) and Terminal-Bench 2.1 (81 vs 71.7).

CICC maintained its "Outperform Industry" rating on Tencent Holdings (00700) with a HKD 666 target price, representing 47% upside from current levels. The broker said the market's expectations for Tencent's AI output are at a low point, and that Hy3's launch — alongside WorkBuddy and WeChat Agent — could drive a valuation re-rating from the current 13x forward P/E.

The reliability pitch that enterprises actually care about

Tencent's internal evaluations show Hy3's hallucination rate dropped to 5.4% from 12.5% in the preview version released in April, while commonsense error rates fell to 12.7% from 25.4%. Multi-turn issue rates improved to 7.9% from 17.4%, and the model's score on the open MRCR long-dialogue benchmark jumped to 75.1% from 42.9%. In WorkBuddy scenarios, task resolution rose to 90% from 72%, with average completion time falling 34%.

These are self-reported numbers, and Tencent's benchmark appendix notes that nearly all competitor scores came from the company's own test runs. Independent verification from indices like Artificial Analysis is still pending. But the choice to foreground reliability metrics over leaderboard scores shows who Tencent is targeting: teams that have been burned by models that demo well and fabricate confidently in production.

Deployment math favors the smaller model

Hy3's FP8 weight footprint of under 300 GB — less than half of GLM-5.2's roughly 744 GB — means a single well-specced node can serve it, versus the 8x H200 node minimum that GLM-5.2 requires. Tencent's recommended serving configuration targets Nvidia's H20-3e, the export-compliant GPU for the Chinese market, but the model runs even more comfortably on H100s, H200s and B200s available in Western data centers through standard vLLM and SGLang deployments.

The pricing reinforces the cost advantage. Tencent charges CNY 1 ($0.15) per million input tokens and CNY 4 ($0.59) per million output tokens, with cached inputs at CNY 0.25 ($0.037). The company claims Hy3 reduces token consumption by 47.4% on document processing and 49% on PPT generation compared with competing models at similar capability levels.

The competitive picture shifts

The Apache 2.0 license — with no regional exclusions or field-of-use restrictions — removes the legal barrier that blocked many Western enterprises from adopting earlier Chinese open-weight models. Tencent says daily token consumption has grown 20-fold since the April preview, with more than 50 internal product teams providing feedback that shaped the final release.

For enterprises weighing open models, the trade-offs are now unusually explicit. GLM-5.2 remains the choice when coding performance is the only criterion and an 8x H200 budget is available. Hy3 makes its case everywhere else: search-heavy agent workloads, reliability-sensitive applications, and organizations that want frontier-adjacent capability without frontier-scale infrastructure. Tencent shares, trading at 13x 2026e Non-IFRS P/E, imply the market has yet to price in any of that potential.

This article is for informational purposes only and does not constitute investment advice.