A 27-billion-parameter AI model normally needs around 54 GB of memory to run. PrismML just shipped one at 3.9 GB, and it runs on an iPhone.
Bonsai 27B, released this week by PrismML, is the first model at this capability tier to clear the memory ceiling of a consumer smartphone. Running on an iPhone 17 Pro Max, it hits 11 tokens per second. The ternary variant, at 5.9 GB, reaches around 26 tokens per second on an M5 Pro laptop. Both are free under Apache 2.0.
Get 1-month GTA+ subscription with pre-order.
Pre-Order GTA 6 Now
How PrismML squeezed 27 billion parameters into a phone
The compression method is built on Caltech intellectual property and works by reducing each model weight from 16 bits of floating-point precision down to a single sign. The binary build uses +1 or -1. The ternary adds a zero state for slightly more expressive range. Every group of 128 weights shares a 16-bit scaling factor, which lands the binary variant at just 1.125 bits per weight, making it 14 times smaller than the full-precision original.
Here's the thing that separates Bonsai from most quantized models: nothing gets a higher-precision escape hatch. Embeddings, attention layers, and the full language model head are all compressed end-to-end. Most low-bit builds keep sensitive layers at full precision to protect output quality, which inflates their file size as a tradeoff. Bonsai skips that entirely.
The model also runs a hybrid attention backbone where roughly 75% of layers use linear rather than full quadratic attention. That design choice is what makes a 262,000-token context window practical on-device hardware, a standard attention stack would make that prohibitively expensive on a phone.
Benchmark performance at 94.6% of full precision
Across 15 benchmarks evaluated in thinking mode on NVIDIA H100 GPUs, covering knowledge, math, coding, and tool use, the ternary variant averages 80.49, which is 94.6% of the full-precision model. The 1-bit build hits 76.11.
For context on what those numbers mean in practice:
- AIME25 and AIME26 math scores: 93.7% for Ternary Bonsai 27B versus 95.3% for Qwen 3.6B
- Coding: 86 points for Bonsai versus 88 for Qwen 3.6
- General knowledge: 77% for Bonsai versus 83% for Qwen 3.6
The key here is that Bonsai achieves those results at a fraction of the file size of comparable models. Conventional 2-bit Qwen builds are nearly twice as large and tend to collapse on math and coding tasks below 4 bits. Bonsai holds together across the board.
Real-world testing: coding and creative writing
The Bonsai team ran the model through a practical test: building a first-person typing-horror browser game called Zombie Type. Two rounds of vibe coding produced clean collision detection, proper scoring logic, and graphics that held together. The model grasps structure early; the second pass refines rather than rebuilds from scratch.
Creative writing is a more qualified story. Zero-shot prompts won't produce anything particularly imaginative. What Bonsai does deliver is consistent internal logic, pacing, and story arc, which puts it on par with or slightly ahead of Claude Haiku on comparable prompts. For a model running entirely on local hardware with no API costs, that's a meaningful result.
PrismML also ships a DSpark speculative decoding layer alongside the model. It works by having a lightweight drafter propose blocks of candidate tokens, which the main model verifies in a single forward pass rather than generating token-by-token. On an H100 that adds a 1.37x throughput boost with no change in output quality. On Apple Silicon it's not yet enabled by default, but it's a real gain for GPU serving.
Apple is paying attention
This is the second major release in the Bonsai family. In March, PrismML shipped Bonsai 8B, a 1.15 GB model that proved the 1-bit architecture could survive at 8 billion parameters without its reasoning falling apart. The jump to 27 billion is where the stakes change.
Apple is now in early talks with PrismML about the underlying compression technology, with the company evaluating it for potential on-device use. PrismML CEO Babak Hassibi confirmed the discussions. A compressed Gemma model is next in the pipeline, followed by larger frontier models.
What most players miss about this announcement is that the implications extend well beyond smartphones. On-device AI at this capability level means no API calls, no latency from server roundtrips, and no data leaving your hardware. For gaming applications specifically, that opens up possibilities for genuinely responsive NPC behavior, local voice processing, and real-time game logic that doesn't depend on a cloud connection.
1-bit Bonsai 27B is available for free download now. If you're looking to keep up with everything in gaming and tech worth following, our gaming guides hub is a solid place to start, and if you're deep into action RPGs right now, the Where Winds Meet beginner's guide and the best mystic arts guide are worth bookmarking.
