AI model competition evolved: some compete on parameters, others compress them into -1 and +1

7.75GB image generation model compressed down to 0.93GB.

Not quantization, not pruning, but directly cutting weights to {-1, +1}.

PrismML recently announced Bonsai Image 4B. This 1-bit quantization reduces weight from 16GB to 0.93GB with minimal loss of quality.

With text encoders and VAEs, the Apple Silicon deployment package fits within 3.42GB—nearly 16GB less than the original. In Ternary mode, weights are 1.125 bits, while in Binary mode, they are 1.71 bits, offering better expressiveness.

Evaluation across three benchmarks confirmed Ternary version retains 95% quality of the original.

"Nothing is noticeable" when reviewing the differences without zooming in—this isn't marketing fluff, but rather a benchmark truth.

Yet I'm more interested in another number

Efficiency is key here: 1.5GB active memory for 9.4 seconds per output image, per authorizing process, result is high quality.

This data translates to:

• Hardware compatibility: iPhone 17 Pro Max capable.
• Real-time testing: "Works on iPhone" becomes "9 seconds for a result immediately after opening the app" on the actual phone.
• App availability: Direct download from App Store as Bonsai Studio.

Unwire.hk tested iPhone Air for continuous generation, noting the phone only became slightly warm after many outputs. Unfortunately, Chinese text support is poor—full-width characters are rendered as garbled text. Sensitivity filters are present; sensitive content is rejected instantly.

A WebGPU demo is also available. Browser-based interface accepts prompts locally, no API key required.

This is what truly defines 1-bit quantization: transforming local image generation from impossible to possible.

Why the "Local" aspect matters

Using Midjourney from cloud requires three minutes per image (1 minute for three iterations). DALL·E still requires paid credits for remote calls with latency constraints.

Image generation is inherently iterative—you modify prompts, adjust seeds, refine parameters, and compare results. On cloud platforms, each iteration involves network round-trip costs and latency. Locally, this workflow is seamless at second-level speed with zero cost.

PrismML's announcement clarifies the choice:

"Cloud APIs are still suitable for many products. However, cloud-only generation imposes certain product constraints: every prompt is a remote request, every iteration carries marginal serving cost, and every interaction adds round-trip latency."

Translation: While cloud APIs are useful, creating a loop where each modification waits for remote servers breaks creativity rhythms and cancels out value.

Getting started in three steps

To avoid theoretical deliberation, proceed directly:

git clone https://github.com/PrismML-Eng/Bonsai-image-demo.git
cd Bonsai-image-demo
./setup.sh

The setup script will download the model."Console formats:

• macOS: MLX format weights
• Linux / Windows: Gemlite format weights

Downloading model versions:

# Recommended: Ternary version for better quality
./scripts/download_model.sh

# For the smallest 1-bit version
./scripts/download_model.sh binary

Generating an image:

./scripts/generate.sh -p "An icy bonsai tree in a rainy forest, photo realistic." --size 1024x1024 --seed 9909

Alternatively, launch web studio:

./scripts/serve.sh
# Open Browser at localhost:3000

Note for Windows users: Update drivers to latest version, or reduce resolution to 512×512 for 4GB or less video cards.

If deployment is unnecessary, check the HuggingFace demo hosted by the Web GPU community.

The competition dimension has changed

A Hacker News discussion with 464 points and over 200 comments left by a veteran community member resonated with me:

"I actually can't wait for the future where I upgrade hardware in order to upgrade my ai as an alternative to an expensive subscription."

This highlights a radical commercial meaning of 1-bit quantization: upgrading AI without paying for a subscription by upgrading hardware.

PrismML calls this "Intelligence Density"—measuring intelligence output per bit, not learning capacity. A 1-Bit Bonsai 8B model requires 1.15GB to run on iPhones, demonstrating competitive performance with 14x larger models. The same logic applies to image generation.

When 0.93GB of models can render images on phones, and 1.15GB language models can compete with 16GB counterparts, does "parameter count" remain the sole standard for AI quality?

Maybe in the near future, models will be compared by another metric: How much efficiency can we achieve with fewer bits?