Bonsai 27B is a July 14, 2026 release from PrismML that re-encodes Alibaba's Qwen3.6-27B into extreme low-bit weights — not a new pretrain. Two variants ship under Apache 2.0: 1-bit binary weights at ~3.9 GB (phone footprint) and ternary {−1,0,+1} weights at ~5.9 GB (laptop quality). Both are multimodal with a 4-bit vision tower, 262K-token context, hybrid attention, and speculative decoding.

Can Bonsai 27B really run on an iPhone?

PrismML targets iPhone 17 Pro / Pro Max class devices: a 12 GB iPhone exposes roughly 6 GB per-app memory, and the 1-bit build at ~3.9 GB is the first 27B-class footprint that leaves headroom for KV cache and activations. Demos use MLX-Swift on device; expect thermal throttling, battery cost (~672 tokens per 1% battery per PrismML whitepaper citations), and cached/prefilled image context in showcase videos — not unlimited cloud-grade agent loops on cellular.

How much intelligence does Bonsai 27B retain vs full Qwen3.6-27B?

PrismML reports ~95% of full-precision baseline on a 15-benchmark thinking-mode suite for ternary (80.5 vs 85.0 overall average) and ~90% for 1-bit (76.1 vs 85.0). Math and coding hold closest; agentic tool-calling and vision drop more on the 1-bit variant. MarkTechPost cites 94.6% and 89.5% retention figures from EvalScope vLLM H100 runs — same ballpark, different averaging method.

How do I download and run Bonsai 27B?

Weights are on Hugging Face (e.g. prism-ml/Bonsai-27B-mlx-1bit). Apple devices use MLX (Python or Swift fork for iPhone/iPad); NVIDIA GPUs use custom CUDA kernels for the hybrid-attention low-bit matmuls. PrismML also offers a limited-time developer preview API. This is not a stock llama.cpp GGUF drop — you need PrismML's Bonsai backends, unlike conventional Qwen3.6 GGUF from Unsloth.

Should I use Bonsai 27B or regular Qwen3.6-27B GGUF?

Use full-precision or Q4/Q8 GGUF Qwen3.6-27B on a 48–64 GB Mac or 24 GB GPU when you want maximum quality and llama.cpp/OpenCode compatibility — see explainx.ai's Qwen 3.6 local guide. Choose Bonsai when footprint is the binding constraint: phone-class 6 GB app budget, single 24 GB GPU serving, or offline agent loops where 90% of 27B beats 100% of 8B. Bonsai is a compression product, not a replacement for cloud frontier models on hard reasoning.

Is Bonsai 27B open source?

Yes — Apache 2.0 license for model weights. PrismML published a technical whitepaper alongside the release. The compression methodology is described as architecture-agnostic, with larger Bonsai models teased in progress.

How does Bonsai 27B compare to Gemma 4 12B 4-bit QAT on a phone?

Hacker News cross-benchmarks (Jul 15, 2026) suggest Bonsai ternary/1-bit crushes Gemma 4 12B QAT (~7 GB) on math and coding, trails slightly on knowledge and tool calling, and falls behind significantly on vision — where Gemma's QAT models are praised for agent speed and pleasant prose. Gemma wins when you need multimodal on-device; Bonsai wins when you need 27B-class reasoning density under 6 GB.

Why does Bonsai fail to load in LM Studio or print exclamation marks?

Bonsai needs PrismML's custom MLX/CUDA kernels or forked llama.cpp — stock LM Studio engines often reject the weights until upstream merges (Jul 15 reports). Android users saw repeated "!" outputs — likely token id 0 in the Qwen tokenizer under aggressive quant or immature backends. Use oMLX, Prism's llama.cpp fork, or Bonsai-demo README paths; expect a few days of integration churn after launch.

← Back to blog

explainx / blog

PrismML Bonsai 27B — 1-Bit 3.9GB Qwen3.6 Model That Runs on iPhone

Jul 14–15, 2026: PrismML Bonsai 27B — 3.9GB 1-bit Qwen3.6 on iPhone. HN thread (502 pts) debates Gemma 4 12B QAT, tool-calling drops, LM Studio load failures, Apple talks. explainx.ai maps benchmarks, harness reality, and honest limits.

Jul 15, 2026·16 min read·Yash Thakker

PrismML