PrismML releases Bonsai 27B, claims first 27B-parameter model to run on-device on iPhone at 4GB memory footprint
PrismML has released Bonsai 27B, claiming it's the first 27-billion parameter model capable of running on-device on iPhone. The model achieves 58-87 tokens per second on Apple's M5 Max chip with a 4GB memory footprint, using 1-bit and ternary quantization to fit within iPhone's approximately 6GB available app memory.
PrismML releases Bonsai 27B, claims first 27B-parameter model to run on-device on iPhone at 4GB memory footprint
PrismML has released Bonsai 27B, claiming it's the first 27-billion parameter model capable of running natively on iPhone. The model achieves approximately 4GB memory footprint through aggressive quantization, fitting within the roughly 6GB of memory iPhones make available to apps.
Performance specifications
According to PrismML, Bonsai 27B reaches:
- M5 Max (Apple Silicon): Up to 87 tokens/second in 1-bit mode, 58 tokens/second in ternary mode
- NVIDIA GeForce RTX 5090: Up to 163 tokens/second in 1-bit mode, 134 tokens/second in ternary mode
The company states the model runs natively on Mac, iPhone, and iPad via MLX, and on NVIDIA GPUs via CUDA through custom low-bit kernels built for its hybrid-attention architecture.
On-device constraints
PrismML explains that running models on iPhone faces stricter constraints than storage alone. A 12GB iPhone provides approximately 6GB of memory available for apps, which must be shared between the model weights, KV cache, and activations. The company claims conventional 27B model builds don't approach this threshold, while the 1-bit Bonsai 27B at approximately 4GB is "the first to pass through with room to work."
Availability and Apple discussions
The model is released under Apache 2.0 License with weights available today. PrismML is offering a limited-time free developer preview API.
PrismML CEO Babak Hassibi told CNBC that Apple and other companies are evaluating the startup's models for speed, energy efficiency, and on-device performance. "They're really evaluating our technology right now," Hassibi said of Apple, characterizing discussions as "very early" with unclear outcomes. Apple has not commented.
The Information first reported last week that PrismML held meetings with Apple about potential uses of its technology.
What this means
If PrismML's claims hold, Bonsai 27B represents a significant advance in model compression for mobile deployment. However, the company provides no benchmark scores for quality or accuracy compared to standard 27B models, making it impossible to assess performance trade-offs from the aggressive quantization. The public discussions about Apple talks appear designed to generate launch buzz—serious partnership negotiations typically remain confidential. Independent testing will be needed to verify both the on-device performance claims and whether the model maintains useful capability at 1-bit and ternary quantization levels.
Related Articles
Meta launches Muse Wonder agent model, gating it behind $20 and $100 monthly Muse app plans
Meta released Muse Wonder, a model it claims is its most advanced for powering digital assistants. Access requires a $20 or $100 monthly Muse app subscription, and a larger Muse Wonder Full variant is limited to Muse Code subscribers. API pricing has not been disclosed.
Odyssey-3 world model enters free public preview; 14B base model, 66.1 Physics-IQ claim carries caveats
Odyssey has opened a free public research preview of Odyssey-3, an autoregressive diffusion transformer that generates interactive environments from text prompts in real time. The base model has 14 billion parameters and outputs 832 × 480 video. Odyssey's headline Physics-IQ score of 66.1 does not meet the benchmark's own requirements for a record claim.
Microsoft releases Decision-1, a Qwen3.5-9B-based model for classification and routing, at $0.042 per 1M input tokens
Microsoft has released Decision-1, a decision model built on Qwen3.5-9B for classification, evaluation, and routing. Microsoft claims 83.5% accuracy across 36 benchmarks and 85 ms latency. Input tokens cost $0.042 per million, and output tokens are free.
Microsoft releases FrogNano-4B, an Apache 2.0 coding agent trained with RL on 1,500 synthetic tasks
Microsoft has released FrogNano-4B-2609, a repository-level coding agent derived from Qwen3.5-4B and published under Apache 2.0 with open weights. Microsoft says it was post-trained only with reinforcement learning on about 1,500 synthetic software-engineering tasks, with no stronger-model trajectories. It is evaluated at roughly 131K tokens of context.
Comments
Loading...