Enterprise AI
PrismML Releases Bonsai 27B as First 27B-Class Multimodal Model Runnable on Phones
The open-weight model derived from Qwen3.6-27B applies 1-bit and ternary quantization to support multimodal reasoning and agentic tasks on mobile hardware under an Apache 2.0 license.
Bonsai 27B is the first 27B-class multimodal model capable of running on a phone.
PrismML announced the availability of Bonsai 27B weights today under an open license that permits commercial and research applications without restriction. The company positions the model as a flagship in its Bonsai family. This open release facilitates integration into various enterprise systems for local AI processing.
The model builds directly on the Qwen3.6-27B architecture which incorporates hybrid attention mechanisms along with a 262000 token context window and an integrated vision component for handling multimodal inputs. This foundation ensures that the quantized versions inherit strong baseline capabilities in reasoning and perception tasks.
What background led to the creation of Bonsai 27B?
Companies in the AI sector have long sought methods to deploy large language models on edge devices such as smartphones to address concerns over data privacy and network dependency in enterprise settings. PrismML addressed this challenge by applying extreme quantization techniques to reduce the memory requirements of a 27 billion parameter model. The effort represents a step toward democratizing access to high capability AI for businesses that require local processing without constant cloud connectivity.
Prior models of similar scale demanded graphics processing units with large memory capacities that are not available in mobile form factors. The focus on 1-bit and ternary representations allows the model to maintain reasoning capabilities while fitting within the constraints of phone hardware. This shift could change how enterprises approach AI deployment strategies by enabling on-device execution for sensitive workflows.
The parent model Qwen3.6-27B provided a strong foundation due to its established performance in complex tasks. PrismML applied custom low-bit kernels to enable native execution on both Apple silicon through the MLX framework and NVIDIA hardware through CUDA. These kernels are optimized for the low precision formats to maximize speed and efficiency during inference.
What new capabilities does Bonsai 27B introduce in detail?
Bonsai 27B extends the Bonsai family with multimodal support that includes vision capabilities alongside text processing. This allows the model to perform tasks that require understanding both images and text in a single inference pipeline. Enterprises can leverage this for applications such as document analysis combined with visual data in operational environments.
The release emphasizes support for agentic loops where the model can call tools and execute multi-step reasoning processes autonomously. Such features are particularly relevant for enterprise applications involving workflow automation. The model can handle sequences of decisions without constant human intervention in routine processes.
Availability under Apache 2.0 means that organizations can integrate the model into their systems without licensing fees or usage restrictions beyond the license terms. This open approach contrasts with proprietary models that limit customization and distribution rights for internal teams.
- Multi-step reasoning processes
- Tool calling mechanisms
- Agentic workflow execution
- Multimodal input understanding
What technical specifics characterize the Bonsai 27B variants?
Two quantized versions are offered. The ternary variant occupies 5.9 GB on disk while the 1-bit version requires only 3.9 GB. These sizes contrast sharply with the full precision version that requires around 54 GB. The compression enables the model to run on devices with typical phone memory configurations that previously could not accommodate models of this parameter count.
The reduction factors enable deployment on devices with limited storage and memory. Custom kernels ensure efficient execution without relying on general purpose frameworks that may not support such low precision. The kernels are designed specifically for the ternary and 1-bit formats to optimize computational throughput on target hardware.
| Variant | Size (GB) | Retention (%) | RTX 5090 (tok/s) | M5 Max (tok/s) |
|---|---|---|---|---|
| 1-bit Bonsai 27B | 3.9 | 89.5-90 | 163 | 87 |
| Ternary Bonsai 27B | 5.9 | 95 | 134 | 58 |
| FP16 Baseline | ~54 | 100 | N/A | N/A |
The architecture retains the hybrid attention from the base model while the quantization is applied to weights to achieve the size savings. This approach preserves the core structure responsible for handling long contexts and multimodal data inputs effectively.
What market and stakeholder implications follow from the Bonsai 27B release?
Enterprise users may benefit from the ability to run sophisticated AI models locally on employee devices. This reduces reliance on cloud services and associated costs while enhancing data security through on-premise inference. Local execution also minimizes latency for real time applications that require immediate responses.
Stakeholders in the mobile device ecosystem including Apple and NVIDIA hardware providers could see increased demand for their platforms as they become viable for advanced AI inference. The open weights allow developers to fine tune the model for specific enterprise use cases such as custom agent behaviors in internal tools.
The license choice supports widespread adoption in regulated industries where model transparency and modification rights are important. Companies can audit the model and adapt it to comply with internal policies without external vendor dependencies.
What expert reactions have accompanied the Bonsai 27B announcement?
Industry observers note the significance of achieving phone compatibility at this scale. The approach demonstrates that aggressive quantization can preserve essential model behaviors even at extreme compression levels. This opens possibilities for broader deployment scenarios.
Today, we're announcing Bonsai 27B, based on Qwen3.6 27B, the new multimodal flagship of the Bonsai family and the first model of its capability class to run on a phone.PrismML
Babak Hassibi, the CEO of PrismML, indicated ongoing evaluations by potential users. This suggests commercial interest in the technology for practical deployments. The statement highlights that enterprises are actively assessing the model for their needs in production environments.
What developments are anticipated next for PrismML and related AI technologies?
Further refinements in quantization methods may lead to even smaller footprints or higher retention rates in future iterations of the Bonsai series. Continued research could push the boundaries of what is possible with sub 4 bit models while expanding supported tasks.
Integration with additional hardware platforms beyond Apple and NVIDIA could expand the reach of such models. PrismML may release additional variants or fine-tuned versions tailored to specific enterprise domains like finance or healthcare.
The success of this release could encourage other organizations to pursue similar strategies for making large models more accessible on consumer and professional devices. This trend may accelerate the shift toward edge AI computing in the coming years.
Frequently asked
How does Bonsai 27B compare to other multimodal models in terms of size and deployability?
Bonsai 27B in its 1-bit form uses 3.9 GB and in ternary form uses 5.9 GB, enabling phone deployment unlike typical full precision 27B models that require around 54 GB.