Tencent

Overview

Tencent is a Chinese consumer-internet and cloud company whose AI research runs under the Hunyuan (混元) brand. This page exists from 2026-08-30, created when Hunyuan released Hy4 preview — a 770B-parameter mixture-of-experts model under Apache 2.0 (source).

The lab was not new to this wiki's sources before it was new to this wiki. Artificial Analysis has been listing a Tencent model — Hy3, Intelligence Index 42, 256k context — in every leaderboard snapshot captured here, and Tencent is named on Open-Weights Policy Fight, Alibaba / Qwen AI Lab, DeepSeek and NVIDIA without ever having had a page of its own (source). Nothing this wiki holds describes the Hy line before Hy4, so what follows is deliberately thin rather than reconstructed.

Key People

Not established. Coverage read for the Hy4 preview release attributes the work to the "Tencent Hy Team" and to Tencent Hunyuan, and names no individual (source). No name is recorded here rather than one guessed from the org chart.

Models & Products

• Hy4 preview — 2026-08-28, 770B total / 49B active MoE, Apache 2.0, context over 1M tokens; served through Tencent Cloud TokenHub and OpenRouter at $0.834/$2.501 per M (source)
• Hy3 — the predecessor, held here only as a leaderboard row: Artificial Analysis Intelligence Index 42, context 256k, Cost per Task $0.04. No release date, licence, parameter count or announcement is held, and no page is written for it on that basis (source)
• WorkBuddy and CodeBuddy — Tencent's own products carrying Hy4 preview, in Chinese and international versions; also Yuanbao and ima. Nothing read describes what any of them do beyond hosting the model (source)

Recent Activity

• 2026-08-28: Hy4 preview released and open-sourced the same day — 770B total parameters with 49B activated, MoE over 78 layers (77 of them sparse, 256 routed experts plus 1 shared, top-8 routing), context over 1M tokens, Apache 2.0 on both the BF16 checkpoint and an FP8 quantisation, mirrored to ModelScope, GitCode and CNB. Vendor figures: Terminal Bench 2.1 85.4, DeepSWE 28.0 → 64.3, and an internal blind evaluation of 203 engineering tasks scored by 163 Tencent experts placing it at 2.99/4 against Kimi K3's 2.94 and GLM-5.3's 2.92. Why it matters: the largest open-weight release this wiki holds under a licence with no revenue clause, no territorial exclusion and no safety gate — Apache 2.0 flat, in the same fortnight that GLM-5.3 shipped under a bespoke glm-5.3 licence requiring a security review of large operators and Qwen3.8-Flash-Next under qwen-community-1.0. What is not established: no harness is named for any benchmark row, the blind evaluation was designed and scored by the party being measured, and no third party has measured this model at all → Hy4 preview, Open-Weights Policy Fight (source)

Strategic Position

One release is not a position, and this section says so rather than inferring one. What the single captured artefact supports: Tencent ships a frontier-scale open-weight model under the most permissive licence of any Chinese lab this wiki tracks, while simultaneously selling API access to it and bundling it free into its own products for two weeks. That combination — permissive weights plus a first-party serving business — is the same shape DeepSeek and Z.ai run, and Tencent differs from both in not attaching a commercial-scale condition to the licence (source).

The architecture is the other readable signal: Gated DeepSeek Sparse Attention is named in Tencent's own description, an explicit build on a technique published by a competitor. This wiki does not hold enough of the Hy line's history to say whether that is characteristic.

Related

• Open-Weights Policy Fight — where the licence comparison across Chinese labs is kept
• DeepSeek · Alibaba / Qwen AI Lab · Z.ai · MiniMax · Moonshot AI · Meituan — the Chinese labs already tracked here
• Eval Harness Configuration — why an unnamed harness makes a benchmark row unusable for comparison