Tencent Hy3 Review: Fast, but not very smart? — Superbash Learn

Superbash Learn

Click to load the YouTube player

Key takeaways

Skip Hy3 for one-shot coding — stick with GLM 5.2 unless you can surround it with MCP servers and external APIs.
Self-host the 295B/21B-active MoE over the cloud API; community sentiment already prefers local inference.
When testing any agent on a known project, `disable code indexing` so the model can't pull references from your other repos.
Free window is `2 weeks` — wire it into Kilo Code, Open Router, or Hermes Agent's News Portal provider before it goes paid.
Treat the roof-scaffolding test as a minimum bar in July 2025: if a model pokes rafters through the top, downgrade it to non-visual tasks.
Run the space-shooter test (enemies must fire at the player) — Hy3 ships a build where targets just drift, so don't ship Hy3-only game prototypes.
Set a hard `~100K-token` budget on Hy3 runs vs. the `200K–300K` you'd give GLM 5.2 — its token efficiency is the one real win here.

Video summary

Companion notes

Hy3 ships as a 295B MoE (21B active) — fast, open-weights, but dumber than GLM 5.2.

Specs and positioning

Tencent dropped Hy3 with 295 billion parameter MoE, 21 billion active parameters, 192 experts, top-8 routing, GQA, a 256K context window, and a 3.8 billion parameter MTP layer for speculative decoding. It's open weights and the community already prefers self-hosting over the cloud API. Tencent's own benchmarks put it slightly below GLM 5.2 across the board against GLM 5.2, DeepSeek V4 Pro, Qwen 3.7 Max, GPT 5.5, and Opus 4.8 — and Boxmining's verdict is blunt: "if they're not stronger than GLM 5.2, then you should just stick to using GLM 5.2."

Speed vs. smarts

Hy3 finished the 3D Chinese architecture project in 3 minutes versus the usual 20 minutes to 1 hour 20 minutes, and burned only ~100,000 tokens vs. the typical 200,000–300,000 for the same build. The medieval castle done in 4 minutes 38 seconds. Helm's Deep in 5 minutes 44 seconds instead of an hour-plus. The space shooter in 2 minutes 40 seconds — a "world record" that comes with the caveat the enemies don't actually fire at the player.

Where it fails

Roof scaffolding was still poking through on the Chinese architecture build, which Boxmining calls a fail since "this wasn't the case 6 months ago" with MiniMax 2.5 / Kimi 2.5. The space shooter "is not a fun game" — enemies don't shoot, the starfield follows the cursor. Helm's Deep "looks nothing like Helm's Deep from the movie"; the gate can be phased through. Logic on every project was weak; visuals were "okay."

How to actually use it

Because raw one-shot capability isn't there (similar to Long Cat), Hy3 needs hand-holding: write "very specific specs," disable code indexing so it can't cheat from existing repos, and bolt on MCP servers and external APIs. Free access for 2 weeks via Open Router, Kilo Code, and News Portal inside Hermes Agent.

Key points

Compare against GLM 5.2 before adopting any new Chinese MoE — Hy3 sits below it on most benchmarks.
Self-host instead of using the cloud API if your box can hold a 295B MoE.
Expect 3–4× the speed of GLM 5.2 and roughly 1/3 the token spend on comparable builds.

Watch on YouTube

Prefer the native player? Open it on YouTube: https://www.youtube.com/watch?v=rkouyv8ea0I