PrismML — 1-bit and Ternary Bonsai Image 4B — local-hardware diffusion image generation

TL;DR

PrismML extends its end-to-end-low-bit Bonsai recipe from LLMs to image generation, announcing 1-bit and Ternary Bonsai Image 4B — a 4B-parameter diffusion image generator targeted at local inference on laptops and phones. The tweet is a launch teaser only: no benchmarks, no model card, no whitepaper link, no architecture or training details. What’s notable is the direction — extending the BitNet-style end-to-end-quantized-weights story from the autoregressive LLM Bonsai release (March 2026) into the diffusion image-generation regime, where most prior low-bit work has been post-training quantization rather than end-to-end training.

Key claims

PrismML is releasing a new model family called Bonsai Image 4B in both 1-bit and ternary weight variants [tweet body].
The stated goal is “high-quality diffusion inference on local hardware: from laptops to phones” — explicitly the same edge-deployment framing as the prior Bonsai LLM release [tweet body].
Parameter count is 4B, matching the mid-size point in the existing Bonsai LLM lineup (8B / 4B / 1.7B) [tweet body].

Method

Not disclosed in the announcement tweet. The HF model cards, runtime forks, and any whitepaper are not linked from this tweet at filing time. Inferences from the prior 1-bit Bonsai 8B release (1-bit Bonsai 8B: End-to-End 1-bit Language Models from PrismML):

PrismML’s prior release used end-to-end 1-bit weights (every projection, MLP, embedding, and LM head at 1-bit), with custom Q1_0 / MLX-1bit kernels in forked llama.cpp and mlx runtimes. Whether the same end-to-end discipline holds for the diffusion U-Net / DiT and its VAE is the central open question for this release.
The “ternary” variant in the LLM lineup uses {-1, 0, 1} weights and is positioned as a higher-quality / less-compressed alternative; this announcement carries both 1-bit and ternary as parallel offerings rather than presenting one as a fallback.

Beyond inference, no claims are made about training compute, dataset, base model, distillation vs from-scratch QAT, sampler, or resolution.

Results

None reported in the tweet.

Why it’s interesting

This is the first datapoint where the BitNet-style end-to-end low-bit weight discipline that’s been the focus of recent LLM-quantization work moves into the diffusion image-generation regime. The prior 1-bit Bonsai 8B release (1-bit Bonsai 8B: End-to-End 1-bit Language Models from PrismML) framed itself as a LLM Inference Efficiency lever orthogonal to speculative decoding and KV-cache reorganization — a lossy weight-compression lever rather than a lossless serving lever. Extending the same recipe to a diffusion model is the harder direction: diffusion inference is dominated by repeated U-Net / DiT forward passes plus VAE decoding, so weight-level quantization affects every denoising step and the VAE separately. The closest filed neighbors in Diffusion serving optimization address the cost in different ways — FlashVSR (FlashVSR v1.1 — Real-Time Diffusion-Based Streaming Video Super-Resolution (HF model release)) compresses temporal redundancy at the architecture/scheduler level, MFLUX (MFLUX — MLX native implementations of state-of-the-art generative image models) ports existing FLUX checkpoints to MLX without changing the weight precision story, and Modal’s “FLUX three times faster” (Run FLUX.1-dev three times faster) is a deployment-stack optimization on top of FP16 weights. Whether end-to-end 1-bit holds quality at diffusion model scale, and how it interacts with the standard CFG / step-count / scheduler axes, is the question the eventual whitepaper or HF model card will need to answer; until then this is a pre-launch teaser that’s worth tracking but not yet a result.