हुनयुआन 3

Tencent Hunyuan's third-generation model: a 295B-parameter Mixture-of-Experts with 21B active parameters routing across 192 experts, built for agent capability and real-world usability. It already powers Tencent products including CodeBuddy and Yuanbao, and Tencent states it matches the intelligence of flagship models two to five times its parameter scale at a fraction of the cost. Unusually, thinking is off by default here and opt-in, over a 256K context.

Hy3 is the general-availability build of a model whose preview had already been in production for ten weeks. Tencent rebuilt the Hunyuan infrastructure in late January 2026, shipped the Hy3 preview in April, then used feedback from more than fifty of its own products to scale up post-training - and it counts that loop, from foundational rebuild to shipped model, as under six months. The lab attributes the improvement to stronger reinforcement learning plus better data quality and diversity rather than to more parameters.

The positioning is unusually practical for a frontier-adjacent release. Tencent names the domains it improved most - software development, office productivity, financial modelling, front-end design and game production - and argues the case with usage rather than scores: average daily token consumption across its products rose twentyfold after the preview shipped, and the number of people actively choosing the model inside its workplace agent grew sixfold. It also ran a blind evaluation with 270 domain experts using tasks from their own jobs, reporting an average of 2.67 out of 4 against 2.51 for the strongest open competitor it tested, with the biggest margins in front-end development, data and storage, and continuous integration.

Much of what changed is reliability rather than capability, and Tencent is specific about it. It reports hallucination rate in its internal real-world evaluation falling from 12.5% to 5.4% and commonsense error rate from 25.4% to 12.7%, under a stated principle of answering when grounded, saying so when the evidence is missing, and never conflating sources. Multi-turn behaviour improved on the operational failures that actually break agents - pronoun resolution, recovering an omitted subject, carrying a constraint forward across turns - with its internal issue rate falling from 17.4% to 7.9%. It also reports that accuracy on SWE-bench Verified varies by less than four points across three different agent scaffolds, which is a claim about portability rather than peak score.

On its own benchmark table Tencent puts Hy3 at 71.7 on Terminal Bench 2.1, 57.9 on SWE-bench Pro, 75.8 on SWE-bench Multilingual, 84.2 on the browsing benchmark BrowseComp, 79.1 on the public MCP Atlas split, 90.4 on GPQA Diamond, and 53.2 on Humanity's Last Exam with tools against 37.0 without them. Where it stops is the frontier: on most of those same rows the lab's table shows the largest closed models ahead of it, and its argument is the ratio, not the ceiling - intelligence comparable to flagship models two to five times its parameter scale, at a fraction of the cost. It is open under Apache 2.0, so the cheapest way to be wrong about it is to not try it.