← Tencent
WorldofAI
Tencent HY3 IS REALLY GOOD! Best Open-Weight Model? (FULLY FREE)
⏱ 15 min video · 3 min read7 Jul 2026
TL;DR
Tencent has officially released HunyuanProver HY3, a 295B parameter mixture-of-experts open-weight model with only 21B active parameters, designed for reasoning, coding, and agentic tasks. The video benchmarks it against DeepSeek V4 Pro, Claude Opus 4.8, GLM 5.2, and Gemini 2.5 Flash across frontend, 3D simulation, SVG, and physics demos, finding it highly competitive at a fraction of the cost.
Key points
1
HY3 is a 295B parameter MoE model with 21B active parameters, 192 experts, top-8 routing, released under Apache 2.0 license — free for commercial use.
2
On SWE-bench Pro, HY3 scores 57.9, beating DeepSeek V4 Pro (55.4), though it trails GLM 5.2 which remains the state-of-the-art open-weight model.
3
In a physics demo comparison, HY3 matched Gemini 2.5 Flash quality across three HTML5 canvas scenes while being approximately 35x cheaper in token cost.
4
HY3 was the fastest model in a multi-prompt coding test (14 minutes) vs Fable 5 (18 min), Sonnet 5 (19 min), and Claude Opus 4.8 (27 min).
5
HY3 is currently free on OpenRouter until July 21, 2026; after that it costs $0.14 per 1M input tokens and $0.58 per 1M output tokens.
Key takeaways
→
Use HY3 now via OpenRouter for free before the promotional period ends on July 21, 2026 — especially for frontend and simulation tasks where it punches above its weight.
→
For pure coding tasks, GLM 5.2 is still the stronger open-weight choice, but HY3 is a strong alternative when visual polish and frontend design quality matter.
→
DeepSeek V4 Pro used the most tokens by far in the physics demo test yet produced the weakest results — HY3 is a more cost-efficient substitute for similar tasks.
→
HY3 supports configurable reasoning effort (no-think, low, high modes), making it flexible for both fast responses and complex multi-step reasoning without always burning tokens.
Notable quotes
“HY3 basically matched Gemini 3.5 across all three scenes and this is around 35 times cheaper.”
“DeepSeek version 4 was the biggest letdown here — it used the most tokens by far and still produced the weakest result.”
“We're getting to a point where we're starting to see results from proprietary giants from these cheap open-source models.”
Worth watching?
⏭️
Worth watching the full video?
The key benchmarks, cost figures, and demo results are all captured here — skip the video unless you want to see the actual rendered frontend and 3D simulation visuals side by side.
Saved you some time? The creator still deserves a like.
Found this by searching? That's the hard way.
summree watches the AI channels you follow and sends you the summary the moment a video drops. Nothing gets past you. From £4/month.
Try it free