Video summary

Qwen3.5 Small Models Compared – 9B vs 4B vs 2B vs 0.8B!

Main summary

Key takeaways

Technology

Summary of technological concepts & product features

  • New “small” Qwen 3.5 models from Alibaba Cloud (dense variants) The video tests multiple released sizes—about 0.8B, 2B, 4B, and 9B parameters—described as multimodal (they can see images and answer based on them).

  • Why they matter (deployment on smaller hardware) Because the parameter counts are low enough, many users can run them on consumer GPUs. The speaker claims the 0.8B model can run even on a recent smartphone.

Key architectural / product traits mentioned

  • All tested variants are dense models (no MoE in this release).
  • Same context length across variants: 262,144 tokens (listed on model cards).
  • Hybrid design for longer context with less VRAM (speaker claims this reduces VRAM footprint).
  • Base model variants are released alongside the main ones for use as starting points for fine-tuning.

Benchmarks availability on Hugging Face

  • Benchmarks are shown only for the 9B and 4B variants.
  • 2B and 0.8B show no visible benchmarks yet; the speaker downplays reliance on these numbers.

Testing toolchain + quantization

  • Inference tool: LM Studio
  • Quantization: Unsloth Q8 for all models
  • Hardware: NVIDIA 590 desktop GPU (provided for the channel)
  • During interactive tests: context length set to 65,536 tokens (speaker corrects an earlier typo about 32,768)

Observed VRAM usage at Q8

  • 9B: ~15.6 GB
  • 4B: ~11.5 GB
  • 2B: ~7.7 GB
  • 0.8B: ~5.8 GB

Sampling parameter control

  • For browser OS / coding-style tasks, models use “suggested sampling parameters” geared toward precise coding.
  • For general multimodal tasks, the speaker switches to suggested general-task sampling parameters.

Main tests performed (reviews / guides / tutorial-style comparisons)

1) Browser OS test (version 2)

Goal: Each model must generate behavior for a “lightweight browser operating system” with:

  • 5 specific apps, including 2 games
  • change wallpaper
  • a special feature chosen and implemented by the model (example: “Zen mode”)

Notable results by model size

  • 9B Browser OS

    • First attempt had an error (developer console indicated a problem).
    • Speaker prompts it like a fix request: “got this error, fix this”.
    • It got stuck in a reasoning loop, recovered by telling it “don’t overthink”; then clock/UI worked.
    • Implemented special feature: Zen mode button.
    • Some UI/game interaction worked only partially (e.g., no right click, limited movement in snake/dodge-style apps), but it produced a desktop that looked functioning overall.

  • 4B Browser OS

    • Faster and appeared more stable on first pass.
    • Special feature: Zen mode.
    • Functional gaps: after closing an app, it couldn’t open it again without refreshing; some games didn’t fully work.

  • 2B Browser OS

    • Speed increases further (about 292.6 tokens/sec).
    • Output was often incomplete or failed to fully “go” in some areas; asking it to fix issues didn’t fully recover results.

  • 0.8B Browser OS

    • Speed: about 396 tokens/sec (very fast).
    • Desktop-like output with buttons and attempted actions (dark mode, wallpaper, calculator), but many actions didn’t actually execute.
    • Even when “fix” attempts failed, output stayed not garbled (speaker considers this a positive sign).

Throughput (tokens/sec, as reported)

  • 9B: ~116.5
  • 4B: ~167
  • 2B: ~292.6
  • 0.8B: ~396

2) Image-to-website multimodal coding tasks (portfolio from wireframe)

Prompt: Provide a hand-drawn/wireframe image and request a beautiful high-tech job-search portfolio website.

Notable results by size

  • 9B: Correct-ish section structure (skills/about/contact) and hover effects, but sometimes content failed to load; included styling such as CRT scanlines and cursor changes.
  • 4B: Best UI transposition into code among the larger models; included moving dots/background and more coherent layout/hover effects, but weaker specificity for TensorFlow/PyTorch initially.
  • 2B: Strong when constrained to “single script using HTML/JS/CSS”; produced more directly AI-relevant competencies (PyTorch/TensorFlow, transformer-related terms).
  • 0.8B: Surprisingly good sub-1B wireframe-to-website conversion, though OCR/name extraction sometimes failed.

Key observation: The speaker emphasizes prompt sensitivity—small models may produce outline-like content instead of code unless explicitly constrained.

3) Virtual drum kit (3D simulation)

Goal: Create a virtual drum kit controllable via keyboard.

  • 9B: Produced a 3D arrangement with propeller-like 3D details, but no sound.
  • 4B: More limited/troubled; partial functionality implied, but not fully working.
  • 2B: UI/keyboard mapping looked better than 4B, but again no sound.
  • 0.8B: Buttons with hover effects, still no sound.

4) “Sven restaurant” website generation (text + aesthetic + photos)

Goal: Generate a website for a fictional restaurant with a “trash becomes luxury / alchemy” theme.

  • 9B: Coherent branding/typography and multiple menu sections with imagery; generally “not bad.”
  • 4B: Strong theme continuity and design, but some garbled/odd sections and missing expected pricing detail.
  • 2B: Creative; typography was difficult to read, but hover-based imagery and menu content existed with variable text quality.
  • 0.8B: Called mind-blowing / a standout, especially for including photos and producing a relatively polished footer and section structure; performance was surprisingly strong for sub-1B.

5) 3D flight simulator from simplified prompt

Goal: Create a simple 3D flight simulator with:

  • a plane
  • keyboard-controlled flight

  • minimal functionality

  • 9B: Most impressive—spinning propeller, camera movement, and flight behavior seemed to work; includes a test for impact/ground reset.

  • 4B: Partially working; complex controls with odd speed/boost/brake behavior and occasional terrain glimpses.
  • 2B: Begins to break down (“wonky”).
  • 0.8B: Minimal, but still produced something resembling the intended behavior.

6) Multimodal “romance novel cover” analysis + chapter outline generation

Prompt: Given a cover image, generate:

  • title
  • logline

  • chapter outline

  • 9B: Lengthy, coherent, emotionally structured outlines with perspective shifts and clear chapter arcs.

  • 4B: Creative and structured, but smaller models “think less” and output may be shorter or less elaborate.
  • 2B: More manageable outlines, still creative.
  • 0.8B: Short outline, but also repetitive.

Notable failure mode (0.8B):

  • When asked to “remove offensive references,” the speaker claims the model produced highly offensive output and appeared to regurgitate or alter references improperly. (The specific policy mechanism isn’t detailed; the key point is inconsistent safety/handling.)

7) OCR capability test using an Arduino/fritzing wiring diagram

Prompt: Identify components in the image and describe what they do.

  • 0.8B: Correct on some major components (e.g., Arduino Uno, breadboard) but mistakes on specific sensor identification.
  • 2B: Better at distinguishing similar sensors; confused “sound level” vs “ultrasonic,” but identified count/function direction more clearly.
  • 4B: Best overall component mapping—identified the motor driver board (even if not the exact model name) and itemized components without duplicates.
  • 9B: Didn’t substantially beat 4B; still missed exact component names (e.g., stepper motor model) though listed sensors similarly.

Takeaway: The speaker suggests 4B hit a favorable accuracy/detail sweet spot for this OCR-like task.

Overall conclusions from the video (comparative “reviews”)

  • General trend: Smaller models run much faster (higher tokens/sec) but show more functional gaps in complex tasks (e.g., Browser OS interactivity and some games/simulations).
  • Best “surprising performer”: 0.8B, especially strong for website generation with photos (Sven restaurant) and decent for multimodal OCR-like tasks.
  • Best “functional” performer among larger models: 9B performed best for tasks needing richer interaction/behavior (browser OS and flight simulator).
  • Task-specific best:
    • 4B: Often strong at component identification and portfolio UI structure.
    • 2B: Good at image-to-website when constrained to a “single HTML/JS/CSS script,” and can include relevant AI/ML competencies.

Main speakers / sources

  • Speaker: Primary presenter/reviewer (name not provided in the subtitles).
  • Model/source: Alibaba Cloud Qwen 3.5 small model family (dense multimodal variants).
  • Platforms/tools: Hugging Face model cards, LM Studio, Unsloth quantization (Q8), and NVIDIA GPU workstation (“Nvidia 590 desktop,” provided by the sponsor/channel).

Original video