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

Product reviewed

Tencent HY3 (HY3 “full release” model). The video frames it as a large Mixture-of-Experts (MoE) model with improved capability versus its earlier preview.

Key features / specs mentioned

• Model size: ~295B parameters (MoE)
  • Active parameters: ~21B
• Performance-related architecture: MTP (multi-token prediction) → expected to run faster than size alone would suggest
• Context length: 256K
• Licensing: confirmed as Apache 2.0 (video notes preview licensing may have been less solid)
• Fine-tuning: includes discussion/resources for fine-tuning, including a fine-tuning guide link
• Quantization expectation: suggested it can run locally when quantized to ~4-bit, fitting on setups like:
  • “two DGX … clusters”
  • 256GB Mac Studio Still “heavy,” but feasible for hobbyist/prosumer with the quantization approach

Major improvements vs preview (claims)

• Benchmarks (JPEGs + stated benchmark comparisons) indicate significant leaps from the preview version.
• Claims that it:
  • outperforms similar-sized models
  • rivals flagship open-source models with 2–5× more parameters
• Community/benchmarks mention it potentially “trades blows” with GLM 5.1/5.2 (per comparisons circulating on X).

Deployment / ecosystem notes

• Initially noted HY3 not showing on OpenRouter, but later concludes it is on OpenRouter and free.
• During the test, the reviewer used a paid route via Zen “Max” / “Claude suggested,” then indicates OpenRouter availability later.

User experience from tests (real-world “chat-to-3D/game” style)

The reviewer runs interactive generation/deployment tests (games/simulations/web/3D), emphasizing whether the model can produce working artifacts, not just visually appealing output.

WebOS 3D / OS-like test

What worked

• Minimalistic interface: clock, start bar, app launching
• Voice control as a “special feature”
• Supports UI tweaks (e.g., custom CSS gradients)
• Basic file explorer and terminal work
• Can run Open GTA (controls mostly work; game runs and shoots)

Issues / limitations

• No right-click
• No “move up a directory” / limited navigation (“can’t get back”)
• GTA variant control oddities:
  • arrow keys not WASD
  • “pedestrian safe version” claim is inconsistent (not fully pedestrian-safe; noted as “interesting” rather than correct)
• Maze 3D: mouse look not working (camera doesn’t move via mouse)

Overall tone: competent but not wow-worthy, and the reviewer expects these kinds of tests may become less relevant over time.

Self-contained C++ “gate” programming test

What worked

• Compiles and runs (after some mode switching and build-step errors)
• Produces a playable “skate” game with generally clean layout
• Output described as surprisingly respectable for the model size—especially that results are “clean” visually and structurally.

Issues

• Second iterative attempt becomes problematic (reviewer closes session; can’t show the result)
• General jank/oddities:
  • invisible rails
  • kickflip orientation issue
  • water/boardwalk movement issues

Overall take

• The reviewer says the first run was better than the second, describing it as “respectable” and “impressive” in context.

Hardware integration: Acer Nitro RGB keyboard under Ubuntu

Goal

• Use the model to create a Linux tool to cycle keyboard RGB colors (keyboard stuck “disgusting orange” under Ubuntu while working in Windows).

What happened

• Model investigated system details (ACPI/protocol assumptions, WMI/control uncertainty).
• Reviewer intervened to correct the approach (avoid “spamming random GitHub repos,” use safer info collection via bash script, handle root access/output).
• Model concluded it couldn’t safely reach per-color commands (EC firmware/internal controls not exposed), but it created a utility to control brightness via what it could access.

Major mishap

• Utility accidentally controlled screen backlight instead of keyboard backlight, making the screen extremely dim, requiring recovery.

Outcome

• Not the requested full RGB cycling, but it demonstrated:
  • reasoning about low-level control paths
  • recovery from a mistake
  • a safe working brightness tool
• Reviewer notes the “entire cost” of this hardware test was $0.43.

Skydiving simulation (comparison with GLM 5.2)

What worked

• Weather/topography/terrain described as really decent
• Controls functional; parachute deployment visually plausible (strings visible; curved parachute look)
• Sound exists
• Score: 96 (“Textbook skydive bullseye”)

Issues

• Some visual oddities in parachute deployment

Comparison

• Reviewer claims terrain quality was better than GLM 5.2, even though GLM 5.2 had more gameplay structure (e.g., rings).

3D city block time travel prompt (comparison with GLM 5.2)

What worked

• Year transitions: 1945 → 2055
• Era styling changes; some hover effects; orbit/camera movement
• Sound cues between eras
• Changing building heights and roof elements
• Reviewer highlights 2055 looks “sick” and futuristic/optimistic

Issues

• Less impressive than GLM 5.2:
  • overall result “less good”
  • cars/pedestrians sometimes incorrect or lacking variety
  • walking animation noted as “a little interesting”
• Reviewer explicitly says it wasn’t at GLM 5.2 level.

Front-end design / website generation

Goal

• Create a Slap Us Watch Co website with:
  • a high-end 3D watch hero section with cinematic panning
  • 3D models for pricing cards

What worked

• “Not too bad” / respectable output for model size
• Scroll effects / fade-in header noted
• Hero watch changes with interaction (e.g., click swaps a watch material/variant)

Issues

• Some weirdness in watch strap
• Orbit/camera height not controllable as expected
• Watch hour tick marks look questionable

V8 engine in OpenSCAD (parametric 3D)

What worked

• Extremely fast generation
• Produced a recognizable V8:
  • “eight individually visible cylinders”
  • crank/intake/manifold-related parts mentioned
• Included appropriate 280 motor mount hole
• When orientation was wrong, the model corrected it quickly

Issues

• Not 100% photorealistic; orientation of some parts likely imperfect (improved quickly)
• Overall: reviewer impressed by speed and functional correctness

Subway-to-FPS with humanoid zombie enemies

What worked

• Produced a functioning FPS conversion conceptually:
  • flickering lights
  • enemies in tunnels
  • brightness control integrated into gameplay
  • no visible weapon recoil (noted as a limitation/difference)
• Playable scenario with “low drama” but “not bad” outcome

Pros (as presented)

• Strong improvements vs preview (claimed significant benchmark gains)
• Strong performance for size due to MoE + active params + MTP
• Works end-to-end in many generation tasks (3D/game/web/parametric CAD)
• Decent visual/terrain quality in simulations (notably skydiving terrain)
• Breadth of capability: from games to websites to hardware-adjacent scripting
• Apache 2.0 license + fine-tuning support
• Potential local feasibility with 4-bit quantization on large consumer/prosumer setups

Cons / limitations

• Some outputs are janky or inconsistent (invisible rails, wrong orientations, animation oddities)
• Hardware control isn’t fully possible for per-color keyboard RGB; at best it reached brightness and caused a screen backlight mistake during testing
• City time-travel task lagged behind GLM 5.2 in quality/variety
• Early ecosystem confusion (OpenRouter availability initially unclear, then confirmed)

Comparisons mentioned

• GLM 5.1 / GLM 5.2 repeatedly used as benchmarks:
  • Skydiving: HY3 terrain better than GLM 5.2 (though GLM had more gameplay elements)
  • City timeline: HY3 noticeably weaker than GLM 5.2 overall
• Community comparisons: HY3 may trade blows with GLM and other flagship open models despite heavier parameter counts.

Ratings / numerical scores

• Skydiving: 96 (“Textbook skydive bullseye”)

Unique points mentioned (consolidated)

• Full release vs preview; benchmark-implied capability jump
• Apache 2.0 licensing (emphasized)
• Fine-tuning resources available
• Architecture/size: 295B MoE, 21B active, MTP for speed
• 256K context length
• Quantization viability: ~4-bit runs on large local systems (256GB Mac Studio, DGX clusters)
• Deployment via providers (Zen “Max” mentioned); OpenRouter initially uncertain then confirmed free
• Interaction/UI notes: voice control + minimal OS UI; no right-click; limited directory navigation
• WebOS games: GTA workable but control quirks; Maze mouse look not working
• C++ gate/skate: compilation mostly succeeds; first run better than second
• Hardware RGB test: per-color cycling not achieved; created brightness utility; accidental screen backlight control; recovered
• Skydiving: good terrain; parachute visuals ok but oddities; score 96
• City time travel: era transitions and effects; 2055 strong; cars/pedestrians issues; worse than GLM 5.2
• Website generation: decent hero/watch rendering and interaction; strap/camera/tick mark issues
• OpenSCAD V8: very fast, functional V8 and 280 motor mounting; orientation correction improved
• Subway FPS: brightness control works; no visible weapon recoil; playable but low drama
• Hardware test cost claimed: $0.43

Speaker breakdown

• Single primary speaker/reviewer runs all tests and provides the main impressions and comparisons (no clearly separated multi-speaker viewpoints).

Verdict / recommendation

Overall: Recommended for users who want a capable, locally-possible (with quantization) multimodel platform for 3D/game/web-style generation.

The video suggests HY3 is competent to strong for its size, shows meaningful gains over preview, and performs well in practical generation tasks—though it doesn’t consistently match GLM 5.2 quality (notably the time-travel city), and hardware RGB per-color control is limited/inaccurate despite some low-level reasoning.