Model Releases
PrismML Releases Bonsai 2 Ternary-Quantized Open-Source Models
PrismML expands its Bonsai series with 2B to 27B ternary-quantized models, enabling high-reasoning performance on consumer hardware via Qwen3.8 lineage.
September 16, 2026
Model ReleaseBonsai 2
Introduction
The barrier between cloud-scale intelligence and local execution is finally crumbling. On September 16, 2026, PrismML announced the release of Bonsai 2, the second generation of its highly successful ternary-quantized model series. While most low-bit quantization methods suffer from catastrophic intelligence collapse, Bonsai 2 manages to preserve complex reasoning and agentic capabilities even in the sub-4-bit regime.
This release isn't just a minor update; it is a massive expansion of the Bonsai ecosystem. Following the success of the original Bonsai 27B in July, PrismML has refined its post-training ternary quantization recipe and extended the family across four distinct sizes: 2B, 4B, 9B, and the flagship 27B. For developers, this means the ability to deploy state-of-the-art reasoning models on everything from high-end workstations to standard laptops and mobile devices.
- Release Date: September 16, 2026
- Model Sizes: 2B, 4B, 9B, 27B
- Quantization Method: Ternary (2-bit class) post-training quantization
- License: Apache-2.0
Key Features & Architecture
The Bonsai 2 family is built upon the proven Qwen3.8 lineages, utilizing a sophisticated post-training ternary quantization process. Unlike traditional quantization that simply rounds weights, PrismML's approach preserves the structural integrity of the model's weights, allowing it to retain the 'thinking' patterns essential for math, coding, and tool calling.
The architecture is designed for maximum hardware efficiency. The flagship 27B model, for instance, is compressed from a ~54 GB FP16 footprint down to just 5.9 GB. This 9.3x reduction allows the model to reside entirely in the VRAM of consumer-grade GPUs or even run at high speeds on Apple Silicon. The efficiency gains extend to throughput, with the 27B model hitting approximately 47 tokens per second on an Apple M5 Max.
- Lineage: Qwen3.8-based architectures
- Efficiency: 9.3x reduction in memory footprint for the 27B variant
- Hardware Optimization: Native support for GGUF (llama.cpp) and MLX (Apple Silicon)
- Scaling: From ultra-light 2B models for mobile to 27B for desktop reasoning
Performance & Benchmarks
The most striking aspect of Bonsai 2 is its intelligence retention. In testing across 14 different thinking-mode benchmarks, the 27B model achieved an average score of 84.78, representing 98.2% of the original FP16 model's intelligence. To put this in perspective, a conventional IQ2_XXS build at a similar memory footprint only scores 72.59, and Bonsai 2 remains within 0.4 points of the much larger UD-Q4_K_XL model.