PrismML's Bonsai 2: Open-Source Ternary Models

Episode 1 · PrismML open-sources ternary Bonsai 2 models, 9B at just 2GB (2026-10-16, 10 posts)

PrismML has released Bonsai 2, an open-source ternary model family available in 2B, 4B, and 9B sizes with file sizes of just 0.43GB, 0.95GB, and 2.0GB respectively. Trained on Qwen3.5, the models target local execution on everyday devices such as phones, PCs, and wearables. The family is live on Hugging Face as the Bonsai 2 Small collection, offering GGUF and MLX 2bit quantized versions that run on WebGPU.

Confirmed

• Model specs: 2B (0.43GB), 4B (0.95GB), 9B (2.0GB), based on Qwen3.5, open weights
• Emphasis on reasoning, visual understanding, and tool-calling capabilities; according to shared posts, instruction-following actually outperforms Qwen3.5
• Available on Hugging Face, including GGUF and MLX 2bit quantized versions, with WebGPU support
• Design philosophy: as @cephaloform relayed HessianFree's summary, on-device models should focus on following instructions, calling the right tools, and staying concise, rather than pushing the ceiling of raw intelligence

Why it matters

Ternary quantization compresses a 9B model to under 2GB, making it realistic to run fairly large models locally on consumer devices like mid-range phones—a representative advance for on-device AI adoption.

• Ternary Bonsai 2 (2B/4B/9B) released: 2B runs in-browser at 500 tok/s on M4 Max — xenovatech · 2026-10-16
• Bonsai 2's philosophy: phone models just need to follow orders, call tools, and stop rambling — cephaloform · 2026-10-16
• PrismML ships Bonsai 2 ternary models: 9B fits in 2GB and beats Qwen3.5 on instruction following — calabi_and_yau · 2026-10-16
• PrismML ships Bonsai 2 ternary models: 9B fits in 2GB for local AI — gajesh · 2026-10-16
• Bonsai 2 Small multimodal models hit HF with 2-bit MLX builds and WebGPU demo — cephaloform · 2026-10-16
• PrismML releases open-weight ternary Bonsai 2 models: 2B at 0.43GB, 90 tok/s on iPhone — gajesh · 2026-10-16
• Prism ML ships ternary Bonsai 2 models (2B/4B/9B) based on Qwen 3.5 — tcarambat · 2026-10-16
• PrismML ships Ternary Bonsai 2 at 2B-27B: 430MB 2B runs 500 tok/s in-browser via WebGPU — victormustar · 2026-10-16
• PrismML's open-source Ternary Bonsai 2 squeezes a 9B VLM into ~2GB — lmoroney · 2026-10-17
• PrismML's ternary Bonsai 2 fits a 9B vision-language model into ~2 GB — lmoroney · 2026-10-17

Episode 2 · Bonsai 2 Tri-value Small Models Tested: 4B Hits the Sweet Spot (2026-10-16, 2 posts)

Hands-on tests of the Bonsai 2 tri-value models (2B/4B/9B) show the 4B is the sweet spot: it runs in just 1GB of VRAM and reaches 80-90% of Qwen3.8 27B's coding and agent benchmarks, while the 27B unexpectedly stumbles on one skill.

• Bonsai 2 ternary models tested: 4B is the sweet spot, but 27B unexpectedly fumbles file writes — cephaloform · 2026-10-16
• Bonsai 2 4B runs on 1GB VRAM, hits 80-90% of Qwen3.8 27B on coding and agentic benchmarks — julianharris · 2026-10-17