Tencent HY3 vs GLM 5.2: The Local AI That Wins at Half the Compute
Tencent just dropped HY3 — a local, open-weight AI model that ties one of its biggest rivals on the hardest science benchmark while burning about HALF the per-token compute. In this deep dive we break down exactly how it works, and how it stacks up against Zhipu's GLM 5.2. HY3 is a sparse Mixture-of-Experts model: 295B total parameters, but only ~21B fire per token (192 experts, top-8 routing). It ships open-weight under Apache 2.0 on Hugging Face, runs on your own hardware — no API, no per-token bill — and packs an 80-layer transformer + a multi-token-prediction layer, grouped- query attention, a 256K context window, and a selectable thinking dial. But "local" has an asterisk... and so does GLM 5.2. We cover the real hardware cost (≈590 GB in BF16, an 8-GPU node), the benchmarks (GPQA Diamond, SWE-bench), the efficiency gap (21B vs 40B active), GLM's million-token context, and finally the verdict: which one should YOU actually run. ⚠️ Note: benchmark figures are approximate / vendor-reported, and some hardware numbers are back-of-envelope calculations, not official specs. Take vendor benchmarks with a grain of salt. ⏱️ CHAPTERS 0:00 A local AI that ties a giant at half the compute 0:22 What "local" really means — open weights, Apache 2.0, your hardware 1:01 Sparse Mixture-of-Experts: 295B total, only 21B active 1:36 Under the hood: 80 layers, MTP, grouped-query attention, 256K context 2:18 The catch: ~590 GB and an 8-GPU H20 node 3:12 Meet the rival: GLM 5.2 (Zhipu, 750B, 1M context) 3:56 The efficiency gap: 21B vs 40B, GPQA & coding benchmarks 5:04 GLM's own asterisk: a million-token context that's mostly theoretical 5:31 The verdict: which model should you run? 6:05 Recap & what's next 🔗 LINKS HY3 (Hugging Face): https://huggingface.co/tencent/Hy3 GLM 5.2 (Hugging Face): https://huggingface.co/zai-org/GLM-5.2 vLLM: https://github.com/vllm-project/vllm SGLang: https://github.com/sgl-project/sglang #LocalAI #HY3 #GLM #MixtureOfExperts #OpenSourceAI #Tencent #LLM #Hunyuan

China Is About To Pop The AI Bubble

AMD Built the DGX Spark Rival I Predicted… But There's a Catch

Tencent HY3 Is VERY Good – Is This a GLM & DeepSeek Competitor?

GLM-5.2: The Complete Guide to the Best Open-Source Model

I Cut My OpenCode Token Usage by 96% - Here's How

open source ai sucks: glm 5.2 versus claude versus chatgpt

I Got DeepSeek V4 Flash Running at 60 tok/s Locally
![Yann LeCun's $1B Bet Against LLMs [Part 1]](https://i.ytimg.com/vi/kYkIdXwW2AE/hqdefault.jpg?sqp=-oaymwEjCNACELwBSFryq4qpAxUIARUAAAAAGAElAADIQj0AgKJDeAE=&rs=AOn4CLDbV4izF3i-wxevCVIn7FJjoy1vlA)
Yann LeCun's $1B Bet Against LLMs [Part 1]

Android 17 sucks. So I put Linux on a phone.

The Golden Age of Second Passports Is Over, Here's What Happens Next

AMD Just Killed AI Subscriptions Forever (Ryzen AI Halo)

Siri AI vs ChatGPT vs Claude in 2026

Mr.Bean Making Celebrities Cry With Laughter NONSTOP!

I Built a FREE AI Trading Bot With Claude + TradingView (Step by Step)

Today's the day...

China Just Started The AI Cold War: Best AI Models Get Locked

Tencent's HY3 Is Free — So Why Can't You Run It?

Creator of C++: Bell Labs, Negative Overhead Abstraction, Mistakes | Bjarne Stroustrup

Revealing The SPECIAL TECHNIQUE Of A Pakistani Man To EXTRACT GOLD From Used Motherboard Waste

