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

• 2026-09-23: A technical report for an 80B open MoE arrived with no number in it, and the model's release date is fifteen months in dispute. Hunyuan-A13B Technical Report — the Hunyuan-A13B Technical Report: Mixture-of-Experts, 80B total parameters activating 13B at inference, pretrained on a rigorously filtered 20T-token corpus with enhanced STEM data curation, post-trained with high-quality SFT plus large-scale reinforcement learning, and carrying a dual-mode Chain-of-Thought framework — fast thinking for routine queries, slow thinking for complex multi-step problems — with high inference throughput claimed for latency-sensitive applications. Performance is given as "competitive across mathematics, science, programming, general language understanding, and agent tasks, often approaching that of much larger models": five domains named, zero scores, no benchmark and no comparator. Why it matters: this page's standing observation is that Tencent is held here largely as leaderboard rows, and that has not changed — the Artificial Analysis table captured 2026-09-27 carries exactly one Tencent row, Hy3 at Intelligence Index 25, 256k context, $0.07 Cost per Task USD, and no Hunyuan-A13B, so the report's claims cannot be checked against the one independent instrument this repo captures. No model page was created and the date is why: the snapshot dates the report 2026-09-23 while every search pass dates the weights to 2025-06-27, fifteen months earlier, and a model page must assert a Released value this wiki cannot support either way. Not established: expert count, routing, context window (search passes give 256K), the RL method, what switches fast thinking to slow, and the licence — search passes give a custom Tencent licence whose territory excludes the European Union, the United Kingdom and South Korea, which is a materially restrictive term for something called open-source and is not read first-party → Open-Weights Policy Fight, Reasoning Models (source) (source)

• 2026-09-01: Two research releases under the Tencent GitHub org, both about what an agent knows rather than what it can generate — ContextPilot: Teaching Agents for Proactive Context Management via Fine-grained RL (github.com/Tencent/ContextPilot, arXiv 2026-08-28) trains an agent to manage its own working context with planning, long-term memory and soft-offloading tools, replacing the trajectory-level reward with action-level advantages estimated by branching at edits flagged by context and entropy variation; reported as stronger and more compact on long-context QA and deep search, with no benchmark, model or figure named in the abstract. The same day's snapshot also carries ElephantBench (2608.28478, github.com/Tencent/ElephantBench), a 1,094-question closed-book probe for long-tail facts on which sources genuinely disagree: across 32 models the strongest recovers both accounts on only 52.4% of questions, and scaling model size and inference-time reasoning improve recall without eliminating the incompleteness. Why it is recorded on this page: everything this wiki holds for Tencent is a generation artefact — Hy4 preview, its quantisation, the Hy3 line — and these are the first entries here from the lab's agent and evaluation research. No author list, affiliation or team is published in anything read, and the org name in a repository URL is the only attribution; arxiv.org answers EGRESS_BLOCKED from this run's sandbox → Agents (LLM Agents), Agentic Reinforcement Learning (source)

• 2026-08-29: A ~200 GiB quantisation shipped one day after the 1.5 TB weights — Hunyuan announced MIX-STQ1_0, compressing Hy4 preview from 1.5 TB to ~200 GiB GGUF through its AngelSlim compression toolkit and publishing it as AngelSlim/Hy4-preview-GGUF. The method is per-layer mixed precision chosen by calibration data rather than a uniform bit-width, ranging from 1.31-bit (STQ1_0) to 2.06-bit (IQ2_XXS). Why it matters: a same-week first-party quantisation is the part of an open-weight release that decides whether anyone outside a datacentre can run it, and it is a step no other Chinese lab in this fortnight's three releases took — Z.ai and Alibaba left community quantisation to third parties. What is not established: Tencent's own quality claim is "it still works well" and no benchmark table, per-task delta or retained-accuracy figure was published; a circulating "about 98% performance" is a Reddit headline, not Tencent's figure, and is not carried here. No hardware target, throughput or memory measurement was published either → Hy4 preview (source)

• 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