I Ran GLM-5.3 Locally For 48 Hours... Local AI is the Frontier! 🤯

Summary of Technological Concepts, Product Features, and Findings

• Local run of GLM-5.3 “full fat” model for 48 hours on the creator’s hardware, targeting a very large (~1.5T parameter-class) model and producing many generations to evaluate performance.
• Benchmark claims vs prior models and competitors:
  • Reported benchmark improvement from GNM/GLM 5.2: 4.6 → 28.3. The creator speculates this may be due to a “max thinking” style evaluation.
  • Claimed the model is beating Kimmy K3 (described as 2.8T params) despite GLM-5.3 being ~700B params (per the creator).
  • Also claims it is strong relative to other frontier/automation-style models (including comparisons mentioning “Plude/Clude…”).
• Licensing note:
  • Mentions a custom license that appears to require contacting the vendor only if revenue is above a threshold—roughly $10B/year (as paraphrased by the creator). Light usage is implied to be allowed.

How the Creator Tests It (Quantization + “Thinking” Modes)

Local model variants tested

1. Standard Q4 (4-bit quantization) — oriented toward speed/feasibility.
2. “INF” edition — treated by the creator as higher fidelity / better quality than Q4.

Key behavioral change: “thinking” cannot be reliably disabled

• For GLM 5.3, the creator says “thinking cannot be turned off” reliably.
• Attempting to disable thinking still leads to behavior the creator describes as “garbage,” while the model starts “thinking” anyway.

Inference speed + token output observations

• Q4 example: ~2,800 tokens for a short “3D solar system” style prompt; low thinking behavior.
• INF example (same style prompt): ~4,000 tokens, ~14 tokens/sec, higher visual correctness.
• Additional runs show very high token counts when configured for more thinking:
  • INF + “high thinking” (extra configuration): up to ~54,000 tokens, slower ~11 tokens/sec
  • “Amusement park” low-thinking run: around 18,000 tokens
  • “Clone of Photoshop” low-thinking run: around 57,000 tokens

Context Attention feature

• The creator enables “context attention,” described as expanding effective extension to tokens during generation.
• Claimed benefits: improved coherence/fidelity for complex outputs (notably the solar system demo).

UI / Interactive Code-Generation Demos (What Worked, What Broke)

3D Solar System demo (UI + interaction)

• Q4: generates a solar system UI, but clicking planets triggers runtime errors (e.g., an “undefined” runtime error).
• INF low thinking: clicking planets works (zoom/navigation), but quality is still limited vs higher thinking.
• INF high thinking + context attention + “system prompts”:
  • Produces a “nicer” and more faithful system; the creator highlights Earth/planets looking more like real continents.
  • Adds a prompt such as “don’t overthink” to manage token/latency.

Spaceship + third-person + lasers

• Q4: spaceship runs visually, but there are bugs/coding omissions (not runtime errors—missing functionality).
• INF: in low thinking, spaceship reportedly works better than earlier models; includes afterburners and laser firing.
• Visual reactions to destruction:
  • The creator reports triggering dramatic effects, including blowing up the sun, then other planets.
• Limitation noted:
  • Gravity/orbits do not change after explosions, and the creator wants gravity to shift too.

Face Generation Interpretation

• Realistic/procedural human face generation wasn’t fully ready for the exact prompt set shown, but the creator presents a prior face result.
• Interpretation: depicts humans as “zombies,” with unusual facial features.
• The creator speculates model naming: “A” might stand for “Alien,” based on the style of output.

Amusement Park / Game UI Generation

• Prompt (from a Reddit-style idea): “Amusement park” with instruction “please don’t overthink.”
• Reported results:
  • Generates a playable amusement park/game UI with interactions (e.g., punching/picking up points/jumping to rides).
  • Multiple bugs observed:
    • NPC interaction issues (punching doesn’t work as expected)
    • Carousel ride moving in the wrong direction
• Creator’s conclusion:
  • Low thinking = simpler/buggier generations
  • High thinking = far more fidelity, reflected in the token count differences between modes.

Photoshop-like UI Generation

• Prompt: “Clone of Photoshop” using low thinking.
• Reported outcomes:
  • Produces a functional UI resembling Photoshop workflows:
    • Layers (adding/toggling/removing)
    • Drawing shapes and applying gradients
  • Limitation: may not open/load real files like actual Photoshop.
• Overall: strong praise for UI polish and usability.

Coding + Correctness Evaluation (Important Technical Takeaways)

Correctness depends heavily on quantization quality

• The creator emphasizes that quantization quality matters for programming/Q&A correctness.

WinRT / product ID case study

• GLM 5.3 Q4 (low thinking): incorrect recommendation—suggests subscription instead of durable.
• GLM 5.3 high thinking: corrects the answer, emphasizing durable, not subscription.
• Warning: Q4 quantization + model choice can produce wrong-but-plausible API guidance.
  • Suggested mitigation: retry/“looping” and/or use better settings.

Additional programming checks

• Swift question: Q4 gets it right.
• C++ question: Q4 gets it right.
• Math question: Q4 gets it right.

Practical recommendation from the creator

• If possible, run a smaller model with higher effective quantization quality, rather than pushing extremely low-bit setups that hurt correctness.
• Expect to do multiple iterations to resolve “nitpicky” errors.

Performance / Feature Roadmap Mentioned

• The creator plans to test MTP (and references prior NTP issues with GLM 5.2), hoping GLM 5.3 will run it better/faster in a future attempt.

Main Speakers / Sources (as identifiable from subtitles)

• Primary speaker: the YouTube creator (speaking in first person throughout).
• Mentioned sources/benchmarked models (not speaking):
  • “Kimmy K3,” “Clude,” “Plude,” “GLM/GNM 5.2,” and “GLM Flash / Quen 4 exp / 3.8 flash” (referenced as comparison points).