When a big company quietly admits it threw away its own AI foundation and started over, that is usually a story of failure. Tencent is telling it as a story of discipline. In mid-October 2026, Tencent Hunyuan (腾讯混元) released and open-sourced Hy3 preview — the first model trained on a rebuilt infrastructure, and by the team's own description the smartest Hunyuan model shipped so far.
The headline number is familiar by now: a 295-billion-parameter mixture-of-experts (混合专家, MoE) model that activates only 21 billion parameters per token. But the more telling detail is the rebuild itself.
What "rebuild" means here
Tencent says it tore down and re-architected the underlying layers — pretraining, reinforcement learning, and infrastructure — rather than patching the old path. For a firm the size of Tencent, rebuilding a training framework means accepting a real risk of short-term performance regression. The payoff they are betting on is a cleaner base for the next several years of models.
Hy3 preview is the first fruit of that bet. Its specs:
- Total / active: 295B parameters, 21B active per forward pass (plus a 3.8B multi-token-prediction layer).
- Context (上下文): 256K tokens.
- Thinking: a fused fast/slow reasoning (推理) mode — the model can switch between quick responses and deeper deliberation.
- Experts: 192 experts with top-8 activation, 80 layers, GQA attention.
The chief AI scientist leading the effort is Yao Shunyu (姚顺雨), who joined Tencent earlier in 2026. He framed Hy3 preview explicitly as "the first step" of the rebuild and said the team is using the open release to collect real feedback from the community before shipping a final version.
Where it is strong
Per Tencent's own GitHub README and the launch coverage, the biggest gains are in coding and agent (智能体) tasks, with solid improvements in complex reasoning, instruction following, and context learning.
- Coding/agent benchmarks: Tencent cites competitive scores on SWE-bench Verified, Terminal-Bench 2.0, BrowseComp, and WideSearch. A Morgan Stanley research note, citing the release, puts SWE-bench Verified at 74.4%, up from 53% on the prior Hunyuan generation and roughly level with Zhipu's GLM-4.7 (73.8%). That specific score is reported via the bank's note, not an independent audit, so treat it as company/bank-cited.
- STEM reasoning: Tencent highlights performance on FrontierScience-Olympiad and IMOAnswerBench, plus results on the Tsinghua Qiuzhen College math PhD qualifying exam (Spring 2026) and the China High School Biology Olympiad (CHSBO 2025).
- Search agents: BrowseComp improved about 40% versus Tencent's own January 2025 baseline.
The throughline is "agentic and multi-step," not "casual chat." Tencent is positioning Hy3 preview as a model that executes inside development environments and open-web search loops — the same framing StepFun and iFlytek are using.
Open weights and where it runs
Hy3 preview's weights and code are published on GitHub, Hugging Face, and ModelScope under an open license, and Tencent documents deployment via vLLM and SGLang. That makes it self-hostable for anyone with the hardware — though 295B total parameters still means a serious rig.
On Tencent's own side, it is already wired into a long list of products: the Yuanbao (元宝) assistant, CodeBuddy, WorkBuddy, QQ, ima, Tencent Docs, and Tencent Lexiang, plus the TokenHub model platform on Tencent Cloud. It also supports popular open agent products such as OpenClaw, OpenCode, and KiloCode.
API pricing on TokenHub starts at 1.2 yuan input and 4 yuan output per million tokens — roughly US$0.17 / HK$1.3 and US$0.56 / HK$4.4 — with an individual developer plan from 28 yuan (≈ US$3.9 / HK$30) per month.
Why a rebuild, why now
The candid answer in the coverage is that Hunyuan had fallen behind the front-runners and needed a new base rather than incremental tuning. A rebuild is expensive and slow, but it removes the compounding debt of an architecture designed for a different era of model scale. Tencent is also explicitly chasing "differentiated model capabilities" through deep co-design with its own products — the model is being shaped by what QQ, Yuanbao, and CodeBuddy actually need.
Trend reading
According to Li An, chief scientist at BrainNet (脑机网), "a rebuild is the most honest signal a large lab can send: it means they judged their old base uncompetitive and chose a harder year over a comfortable decline."
Honest limitations
- The 74.4% SWE-bench Verified figure comes from a Morgan Stanley note citing Tencent's release; we have not independently reproduced it, and benchmark scores from any single source should be treated as indicative, not authoritative.
- Hy3 preview is, as the name says, a preview. Tencent has not published a final license or stability guarantees, and weights can still change before the formal release.
- "Fast/slow thinking fusion" is a capability claim; how much the slow mode costs in latency and extra output tokens depends on your configuration.
- 295B total parameters makes self-hosting expensive despite the lean 21B active count — open weights do not mean light hardware.
- The rebuild narrative is Tencent's own framing; outside observers cannot directly verify how much of the old stack was truly discarded versus reused.
What readers can do now
- Pull the Hugging Face weights and run a single long agent task (a multi-file refactor, say) to judge the fast/slow mode trade-off on your own hardware.
- If you already use Tencent products, test Hy3 preview inside CodeBuddy or Yuanbao before assuming the open-weight build behaves identically to the hosted one.
- Track the formal (non-Preview) release and the license terms before committing it to a commercial pipeline — preview checkpoints are not production contracts.