Although there are many small LLM models designed specifically for mobile phones, everyone who has used them knows that because they have too few parameters, they are not very easy to use or smart, and are completely incomparable with many online AI services. However, this is due to hardware limitations. The memory space of mobile phones is limited, and it is impossible to install smart enough AI models. For example, if you want to use a super powerful brain with 27 billion parameters (27B level) on your mobile phone to handle complex tasks, the only way in the past was to connect to the Internet.
And that may be starting to change. Foreign AI startup PrismML recently launched the new Bonsai 27B model, which successfully compressed the 27.8 billion parameter model that originally required about 54GB of memory to run to only 3.9GB. PrismML claims that this is the first 27B level model that can fit into the memory space of a mobile phone, and it can already be executed directly on the iPhone.
PrismML launches Bonsai 27B: 54GB model compressed to 3.9GB, 27 billion parameter AI can run natively on iPhone
Bonsai 27B is a multi-modal model based on Qwen3.6 27B. It can process text, images, screenshots and documents, and also supports reasoning, programming, tool calling and AI agent work.
This time, PrismML has launched two versions: the 1-bit Bonsai 27B, which focuses on extreme compression and is prepared for mobile phones, and the Ternary Bonsai 27B, which is slightly larger and prioritizes retaining model capabilities. The focus of this article is the previous 1-bit Bonsai 27B version:
The original Bonsai 27B model has a total of 27.8 billion parameters. If each weight is stored in the common FP16 format, the light model itself requires about 54GB of memory. Even if pressed into a normal 4-bit version, it still takes up about 18GB, which is still too large for mobile phones and most laptops.
What PrismML does is to greatly simplify the values that can be used for each weight in the model. The weights of general models can record many different values, but the 1-bit Bonsai 27B mainly only allows the weight to choose between `−1` and `+1`, and then uses FP16 group scaling values, so that each weight occupies approximately 1.125 bits.
If you want to explain it in a more vernacular way, you can think of the original model as a 4K photo with high image quality, rich colors and light and dark levels. 1-bit compression is like converting it into a black and white silhouette. Although a lot of colors and details are missing, as long as the important outlines are preserved, we can still recognize who the person in the photo is. What PrismML tries to retain is the most important “outline” of the model for mathematical, logical and programmatic reasoning.
The final weight file of 1-bit Bonsai 27B is about 3.9GB, which is about 14 times smaller than the 54GB FP16 version, and the volume is reduced by more than 92%.
As for the Ternary Bonsai 27B mentioned earlier, the file size is about 5.9GB, and the weight can use three values: `−1, 0, and +1`. It sacrifices a little size in exchange for quality closer to the original, which is more suitable for laptops.
PrismML also announced 15 benchmark tests covering mathematics, programming, knowledge, instruction following, tool calling and visual processing. Judging from the official classification average scores, the overall average score of 1-bit Bonsai 27B is 76.1 points, and the original Qwen3.6 27B is 85.0 points, which is equivalent to retaining about 90% of the average performance.
| AI capability metrics | Original Qwen3.6 27B | 1-bit Bonsai 27B | What does it mean to the average user? |
|---|---|---|---|
| mathematical reasoning | 95.3 | 91.7 | Hardly dumber.Arithmetic and logical reasoning are among the abilities that remain most intact after compression. |
| Programming | 88.7 | 81.9 | Still useful.Writing simple scripts, spreadsheet formulas, and assisting with debugging are still useful. |
| Tool calls and agent work | 80.0 | 66.0 | Obviously struggling.The longer a multi-step operation is, the higher the chance of errors during the process. |
| instructions to follow | 78.4 | 65.8 | Complex rules are easy to miss.When there are many formats or conditions, stability will be affected. |
| knowledge and science | 83.1 | 73.4 | General Q&A available.But it cannot be directly used as a replacement for the cloud flagship model. |
| visual processing | 72.6 | 59.6 | A bit presbyopic.The ability to read pictures, identify documents and understand images is the most discounted. |
| Overall average of 15 tests | 85.0 | 76.1 | Retains about 90% of the official average performance.With a size reduction of over 92%, the swap ratio is attractive. |
From the above tests, we can see that if you only use the 1-bit Bonsai 27B to chat, organize data, calculate, or write simple programs, the impact of compression may not be too obvious. But if you want it to view images or call tools and agents, such as automatically opening the app, searching for restaurants, filling in information and completing reservations, the stability will be very different from the original version.
In terms of speed, PrismML claims that 1-bit Bonsai 27B can reach about 11 Tokens per second on iPhone 17 Pro. Taking local chat as an example, 11 Tokens/second can already continuously generate content, so you won’t have to wait a long time for every question.
The same model was also tested on the M5 Max with stronger hardware. The official data can reach up to 87 Tokens per second; with NVIDIA GeForce RTX 5090, it can reach up to 163 Tokens per second.
Running on M5 Max:
Running on an RTX 5090 computer:
However, 3.9GB refers to the model weight size, not the complete memory requirements of the entire model after it starts chatting. PrismML’s GitHub description shows that the peak memory requirement of the MLX 1-bit version is about 5.5GiB after adding the computing element, KV Cache and the context of about 4K Tokens. If you increase the context to 10K, it comes to about 5.9GiB.
In addition, according to foreign media reports, Apple has begun to contact PrismML. PrismML CEO Babak Hassibi said Apple and other companies are evaluating their models to measure device-side speed, energy efficiency and real-world performance. He described the discussions as “very early days” and it was unclear where they would end up.
Bonsai 27B was released on July 14, 2026. The model weights are licensed under Apache 2.0 and can be obtained from Hugging Face download. The 1-bit and Ternary versions can be executed on Mac, iPhone and iPad through MLX, and also support CUDA of NVIDIA GPU, as well as llama.cpp’s CPU, Metal, Vulkan and other backends.PrismML currently also provides online experience, interested people can try it out.
Source: KOCPC Chinese