I Ran Hy4 770B Locally for 48 Hours... Local AI is the Frontier! 🤯

Technology/Model Being Tested

• The video benchmarks Tencent’s “Hi4” model (title implies HY4 / Hi4 770B), with emphasis on running it locally for extended testing (48 hours).
• The host notes it’s a preview/preliminary version.
• Reported performance claims:
  • SWE-65 reportedly beats DeepSeek V4 Pro, Qwen 3.0, and GLM 5.3.
  • Rankings can vary slightly depending on the dataset; GLM is sometimes ranked slightly higher.

Architecture / Training Signals (as described)

• The model is described as assembling “best solutions” from other models by borrowing components/approaches, including:
  • DeepSeek V4
  • DeepSeek V3 (3.2)
  • A “Kimmy Delta retention mechanism”
• Index optimization is said to be borrowed from GLM 5.3.
• The host frames the model’s code and architecture as “fascinating,” and suggests this contributes to strong local generation behavior.

Local Deployment Setup

• Two locally-run quantized variants are tested:
  1. Standard quantized Q4
  2. “INF A” quantized release (described as optimized and better-performing)
• Hardware mentioned for local runs:
  • Apple M3 Ultra / M5 Max / Mac Studio-class systems (performance varies by machine)
• Resource usage (as stated):
  • Q4: ~404 GB
  • INF A: ~424 GB, still within a 512 GB limit
• Throughput and session length:
  • With “thinking” enabled, the host reports outputs frequently reaching ~9.5–9.8 tokens/sec
  • Some generations run for very long sessions (tens of thousands of tokens), implying substantial compute time

Key Feature: “Thinking” Toggle

A central point is a switch that enables/disables “thinking” during generation.

• Observed behavior with Q4:
  • Enabling “thinking” can cause loops/repetition
  • Or it can trigger runtime failures in some interactive/code-like tasks
• Observed behavior with INF A:
  • “Thinking on” tends to:
    • reduce hanging (e.g., solar system lighting works)
    • improve 3D/interactive generation quality
  • Still may produce artifacts, and can result in compilation/shader/runtime errors
• Comparison to GLM 5.3:
  • The host claims GLM 5.3 cannot disable thinking, while Hi4/Hi4 770B allows it.

Tutorial/Test Results by Category (main review content)

1) 3D Procedural Content: Solar System

• Q4 (thinking off):
  • Works, but planets lack lighting/backlighting
  • When load increases and “think” is enabled, it can repeat in a loop
• INF A (thinking on):
  • Produces a solar system where all planets are illuminated
  • The host reports no compilation/runtime errors and calls it a “victory”

2) Image/Art/Procedural Faces

• Procedural face generation works well, especially cosmetically
• With “thinking on,” the host expects more detail (better lighting/structure)
• Earlier outputs reportedly show oddities typical of quantized runs (e.g., limitations such as missing or incorrect upper/lower lip behavior)

3) Interactive Code/3D Editor Behavior (“runs” vs runtime errors)

• The host evaluates whether generated code is:
  • working
  • not stuck in endless bug-fixing
• Reported findings:
  • Under some conditions (certain “thinking” levels), the model may produce runtime errors or UI interaction issues
• Overall claim:
  • The host says many generations succeed without errors, positioning this as an advantage

4) Game/Simulation Generation (Canvas/3D Levels)

The host runs multiple mini “games”/levels, comparing thinking off vs on, and Q4 vs INF A.

• Racing / Motorcycle

  • Q4: works but slower, with small bugs; may require multiple runs
  • INF A: more complete driving experience (steering/combos), with some artifacts at medium thinking but fewer hard failures

• Spaceship / Lasers

  • Thinking off: sometimes shaky behavior or runtime errors
  • Thinking on: more sustained/complex code, but takes longer

• Red Dead Redemption-style

  • Thinking off: simplified environment and basic combat interactions
  • Thinking on: more complete scenes/cutscenes
  • Host notes long compute time (thousands of tokens → long wall time)

• Chaos Park Simulator (theme park level)

  • Thinking off: partial/early “works but error-prone” behavior
  • Thinking on: major upgrade—roller coaster, people, security interactions, throwing pies, water pistol, stink bombs, etc.
  • Framed as a “fully finished theme park level,” though ride controls/entrance may be unclear

• Karting

  • Thinking on: more complete physics/interaction demo (collisions, turbo boost, steering)
  • Host compares it to results from DeepSeek V4 Vision

• Procedural City

  • Thinking off: city present but less coherent (e.g., cars go wrong directions)
  • Thinking on: larger and more detailed city with neighborhoods and many landmarks:
    • bus station, school, stadium, business district, airport/airplanes, shopping mall, etc.
  • Host also notes character interaction: ability to select characters and control them

• Twilight Forest (FPS horror shooter with zombies)

  • Thinking on: described as ~50,000 tokens generation
  • “No sound” mentioned; host plays remotely and notes cursor-control difficulty
  • Reports multiple enemy waves, dark atmosphere, and strong enemy pressure (“good psychopath skills”)

• Asteroids

  • Early outputs are very basic
  • Some runtime issues occur (e.g., failed to create shader)

5) Quantization Stress Test: Animation/Rendering Artifacts

• The host checks quantization quality by generating an animation (car driving)
• Typical artifacts when quantization is bad:
  • misaligned wheels/windows
  • incorrect rotation/direction
  • headlights pointing the wrong way
• The host presents INF as passing this test better than expected (no direct Q4 vs INF comparison is shown for this specific demo)

6) Office-like Tool Generation (MS Word mention)

• With “thinking off,” the model allegedly attempts to open MS Word, but sometimes fails or triggers runtime errors in interaction-heavy scenarios

Product/License and Commercial Use Claim

• The host emphasizes the model has a full Apache 2.0 license.
• The host claims this license is more permissive than some API/model providers that impose additional usage obligations (e.g., contacting providers if earnings exceed a threshold).

Overall Conclusions (as presented)

• INF A quantized variant is generally favored:
  • better stability with “thinking”
  • stronger interactive/3D lighting and scene coherence
  • still subject to quantization limits (some missing facial features, occasional runtime/shader errors)
• The thinking toggle is repeatedly emphasized as crucial for balancing:
  • output quality
  • runtime stability
  • compute time
• The host suggests future updates (including a K3 update / merged version), and notes tests vs cloud/federated variants sometimes produced better results.

Main Speakers/Sources

• Primary source/speaker: The video host (single narrator/reviewer)
  • Runs HY4/Hi4 locally
  • Compares Q4 vs INF A
  • Evaluates the impact of “thinking” settings