Tencent Hy3 Review: Fast, but not very smart?
TL;DR
- Hy3’s clear win was throughput: Ron’s four project runs finished in 2:40 to 5:44, far faster than the comparison times he normally saw for these projects. (source video rkouyv8ea0I, 03:15; 05:22; 07:43; 09:10)
- Speed did not translate into dependable results. The Chinese roof kept exposed scaffolding, the castle had structural logic errors, the shooter enemies did not fire, and Helm’s Deep did not resemble the movie location. (source video rkouyv8ea0I, 02:48; 04:00; 07:21; 08:17)
- The first architecture run used roughly 100,000 tokens, versus Ron’s usual estimate of 200,000–300,000 for that project. That is one recorded test, not a universal cost guarantee. (source video rkouyv8ea0I, 03:38; 03:32)
- Ron’s practical route is to write more specific specifications and give Hy3 more tools, MCP servers, or APIs. For raw one-shot coding with limited tools, his verdict is that it is not there yet. (source video rkouyv8ea0I, 05:50; 06:31; 06:49)
Ron’s verdict
Hy3 is a fast executor that needs supervision, not the model Ron would trust to interpret a broad creative brief and ship a polished result in one pass. Its low run times and the first run’s token use make it interesting for tightly scoped work, but each demo exposed missing logic, weak visual fidelity, or both. If you can write an exact spec, add the tools the task needs, and inspect the output, test it. (source video rkouyv8ea0I, 03:38; 05:50; 06:41) If you need strong raw one-shot capability—finishing a prompt correctly without a repair loop—Ron says to stay with GLM 5.2 or a stronger option from his tests. (source video rkouyv8ea0I, 01:55; 06:49; 09:23)
Key moments
- 00:00 — What Hy3 is: Ron introduces Tencent’s open-weights model and the free-access window announced in the video.
- 00:29 — Specs and GLM 5.2 comparison: the model’s size, routing, context, speed claim, and benchmark positioning.
- 02:15 — Chinese architecture test: fast completion, low token use, and an unfinished roof.
- 03:43 — Medieval castle test: better detail, but questionable construction logic.
- 07:00 — Space shooter test: a record-fast run produces a game without working enemy attacks.
- 08:05 — Helm’s Deep test: self-correction during reasoning cannot rescue scene fidelity or collision logic.
- 09:48 — Final routing advice: try the launch offer, add tools, and do not expect one-shot quality.
Useful quotes
“I think it’s definitely faster, but not maybe smarter.” — Ron, source video rkouyv8ea0I, 00:42
“So, the logic not good, visuals is okay.” — Ron, source video rkouyv8ea0I, 04:11
“This is not a fun game.” — Ron, source video rkouyv8ea0I, 07:41
“Definitely needs more help, that’s for sure, but one shotting, not there.” — Ron, source video rkouyv8ea0I, 10:10
What the four tests actually showed
The video describes Hy3 as an open-weights model that can be self-hosted, with 295 billion total parameters, 21 billion active parameters, and a 256K context window. Ron also says the community sentiment he saw favored local use over the cloud API. These are specifications and observations reported in the video; this companion does not independently verify them. (source video rkouyv8ea0I, 00:30; 00:52; 01:04; 01:10)
The important evidence is not the spec sheet. It is the same pattern across four projects: Hy3 gets to an output quickly, then misses constraints that determine whether that output is usable.
| Test | Speed and efficiency recorded by Ron | Quality check | Video result |
|---|---|---|---|
| 3D Chinese architecture | 3 minutes and about 100,000 tokens; Ron says comparable runs usually take at least 20 minutes and roughly 200,000–300,000 tokens. (source video rkouyv8ea0I, 03:15; 03:19; 03:38; 03:32) | The roof completed, but sticks or scaffolding still poked through instead of producing the polished roof Ron expected by July. (source video rkouyv8ea0I, 02:48; 03:05) | Fail. (source video rkouyv8ea0I, 03:05) |
| Medieval castle | 4 minutes 38 seconds; Ron expected about 20–30 minutes. (source video rkouyv8ea0I, 05:22; 05:26) | Time-of-day shadows, siege x-ray, stairs, and keep detail impressed him. Floating structure and build order did not make sense. (source video rkouyv8ea0I, 04:00; 04:18; 04:48) | Visual pass, weak logic. (source video rkouyv8ea0I, 05:08; 05:18) |
| Space shooter | 2 minutes 40 seconds. (source video rkouyv8ea0I, 07:43) | The starfield followed the cursor, firing was automatic, and enemies did not shoot back—the core interaction Ron expected. (source video rkouyv8ea0I, 07:09; 07:21; 07:35) | Big fail. (source video rkouyv8ea0I, 07:24; 07:32) |
| Helm’s Deep | 5 minutes 44 seconds; Ron says a project at this scale would usually take at least an hour. (source video rkouyv8ea0I, 09:10; 09:13) | Rain worked and the agent corrected some code mid-reasoning, but the scene did not resemble the movie location and Ron could phase through the gate. (source video rkouyv8ea0I, 08:17; 08:38; 08:54; 09:06) | One-shot capability not there. (source video rkouyv8ea0I, 09:23) |
Companion analysis: The architecture test gives the clearest efficiency signal because Ron supplies both time and token comparisons. The other runs support the speed pattern, but the video does not provide token totals for them. None of the four demos proves API cost, production reliability, or performance on non-visual coding tasks.
How to test Hy3 without fooling yourself
Ron disabled Kilo Code’s code indexing so the agent could not pull examples from his other repositories. He also denied external APIs, build tools, and NPM to expose the model’s raw capability. (source video rkouyv8ea0I, 06:17; 06:35) Companion analysis: This isolates raw capability more aggressively than a tool-rich production setup would.
Use this decision aid:
- Do you need a polished one-shot creative build? → Skip Hy3 on this evidence. Every test needed more logic or visual correction. (source video rkouyv8ea0I, 03:05; 05:08; 07:32; 09:23)
- Can you provide a very specific spec? → Hy3 becomes more plausible. That is Ron’s main prompting advice after these tests. (source video rkouyv8ea0I, 05:50; 06:31)
- Can you add MCP servers or external APIs? → Test a tool-assisted run rather than assuming the limited-tool result is the ceiling. Ron proposes those additions but does not show a completed retest, so the improvement remains unproven here. (source video rkouyv8ea0I, 06:41; 09:18)
- Are you comparing models fairly? → Disable code indexing when measuring raw capability, then run a separate real-work test with the tools you would actually allow. The first control comes from Ron’s setup; the second run is companion guidance. (source video rkouyv8ea0I, 06:17)
What changed since this video
The video was published July 7, 2026, and this companion was source-checked on July 18, 2026 against the saved transcript and timestamp segments. No outside release note, pricing page, benchmark, model card, or follow-up test was added. The “next 2 weeks” free access through OpenRouter, Kilo Code, and the provider route inside Hermes Agent is therefore preserved only as a launch-time statement from the video, not a claim that the offer or integrations are available now. (source video rkouyv8ea0I, 00:05; 09:48; 09:56)
Related
Watch on YouTube
Prefer the native player? Open it on YouTube: https://www.youtube.com/watch?v=rkouyv8ea0I