BONSAI 27B · JUL 14, 2026 · 3 MIN READ

Can Your Phone Run Bonsai 27B? RAM Requirements & Speed Estimates (July 2026)

3.8 GB — 1-bit build~11 tokens/s on iPhone 17 Pro Max27.3B params
bonsai-27bprismmlon-devicerequirementsupdated

Quick answers:

  • Yes, flagship phones can run Bonsai 27B — the 1-bit variant needs ~5.4GB of free RAM (3.8GB file + overhead), the ternary variant ~9.2GB.
  • 1-bit (3.8GB file): runs on 12GB+ flagships comfortably, and on 8GB phones only with a small context window.
  • Ternary (7.2GB file): needs a 16GB Android; 12GB phones are borderline. (PrismML's "5.9GB" figure is the theoretical weight size — the actual GGUF on Hugging Face is 7.17GB.)
  • Expect roughly 8-14 tokens/s on 2024-25 flagships — PrismML reports ~11 tokens/s on iPhone 17 Pro Max.
  • It's multimodal (4-bit vision tower adds ~0.5GB if enabled).

If you search for bonsai 27b requirements, bonsai 27b on android, or can my phone run bonsai 27b — this is the reference table.

What is Bonsai 27B?

Released July 14, 2026 by PrismML, Bonsai 27B is a 27.3B-parameter model distilled from Qwen3.6-27B, quantized end-to-end to 1-bit (1.125 effective bits/weight) or ternary (1.71 bits/weight). PrismML reports the 1-bit build retains ~90% of full-precision performance across 15 benchmarks. Context window: 262K (hybrid attention keeps the KV cache small). It is the first 27B-class model that genuinely fits phone memory budgets.

RAM math

Variant File size (HF actual) + runtime & KV overhead* Free RAM needed Verdict
1-bit (Q1_0, 1.125bpw) 3.8 GB ~1.6 GB ~5.4 GB 12GB Android / iPhone 17 Pro: ✅ · 8GB phones: ⚠️ tight
Ternary (Q2_0, 1.71bpw) 7.2 GB ~2.0 GB ~9.2 GB 16GB Android: ✅ · 12GB phones: ⚠️ borderline · 8GB: ❌

*4k context. Bonsai's hybrid linear attention keeps KV small; long contexts (32k+) add roughly 0.5-1GB more.

Which phones run it? (by device)

Phone RAM 1-bit (3.8GB) Ternary (7.2GB) Est. speed†
ROG Phone 9 Pro / OnePlus 13 (24GB) 24GB ✅ Easy ✅ Easy ~9-12 tokens/s
Galaxy S26 Ultra / OnePlus 15 (16GB) 16GB ~10-13 tokens/s
Galaxy S25 Ultra / Xiaomi 15 Pro 16GB ~9-12 tokens/s
Galaxy S24 Ultra 12GB ⚠️ borderline ~8-11 tokens/s
Pixel 9 Pro 16GB ~8-11 tokens/s
iPhone 17 Pro / Pro Max 12GB ❌ (iOS per-app limit) ~11 tokens/s (PrismML)
iPhone 16 Pro 8GB ⚠️ small context only ~7-9 tokens/s
Galaxy S24 / most 8GB Androids 8GB ⚠️ tight ~7-9 tokens/s
6GB phones (iPhone 15, budget Android) 6GB

†Formula estimates from memory bandwidth; will be replaced with measured numbers as we test devices. PrismML's own figure for iPhone 17 Pro Max is ~11 tokens/s (1-bit).

How to run it

  1. Get a GGUF build: prism-ml/Bonsai-27B-gguf (1-bit) or prism-ml/Ternary-Bonsai-27B-gguf on Hugging Face; Apple-silicon MLX builds also exist.
  2. On Android/iOS: load it in PocketPal (or ChatterUI on Android) as a local GGUF model.
  3. Start with 4k context. If the app crashes on load, your phone doesn't have enough free RAM — reboot, or fall back to a 4B-8B model.

FAQ

Does Bonsai 27B beat a 7-8B model on a phone? PrismML's benchmarks say yes at ~90% of the full 27B — but 1-bit quantization quality is workload-dependent. For coding, compare against Qwen3 8B Q4 before committing to the 4GB download.

Can it see images? Yes — the vision tower ships 4-bit (~0.5GB extra when enabled).

Why does it fit when Qwen3 32B doesn't? Bits per weight: a Q4 32B needs ~18GB; 1-bit 27B needs 3.8GB. Same class, 4-5× smaller footprint.


Estimates use our standard fit methodology; measured entries will be labeled separately as evidence arrives. Data sources: PrismML announcement and documentation, plus public repository file sizes (accessed 2026-07-17). Try the phone checker for a specific device.