PrismML released Bonsai 2 27B on Sept. 17, compressing a 27-billion-parameter multimodal model into a 5.9-gigabyte footprint designed to run on local hardware rather than depend entirely on cloud infrastructure.

The Pasadena, California, startup based the model on Qwen3.8 27B and released its weights under the Apache 2.0 license. TechCrunch independently confirmed the launch and reported that the model is part of PrismML’s effort to make capable reasoning systems practical on personal computers and smartphones.

PrismML claims near-full capability in a smaller model

In its release announcement, PrismML said Bonsai 2 uses ternary weights and requires more than nine times less memory than the full-precision model. The company reported that it retained 98.2% of aggregate benchmark performance across tests covering reasoning, mathematics, coding, instruction following, vision and agentic tool use.

Those are PrismML’s own benchmark results and have not been independently validated. The company also reported throughput of as much as 143 tokens per second on an Nvidia GeForce RTX 5090 and 46.8 tokens per second on an Apple M5 Max. Results will vary by hardware, workload and implementation.

The model supports text and image inputs and a 262,000-token context window. PrismML said it runs on Nvidia GPUs through CUDA and on Apple devices through MLX. The company’s public demonstration repository provides code and model links for developers.

Local inference changes the deployment calculation

A smaller memory footprint can let companies run more AI tasks on devices or workstations, reducing network latency and limiting how often sensitive data must be sent to an outside cloud. It can also create a hybrid architecture in which local models handle frequent or private work and cloud models receive only the tasks that need greater capacity.

The release does not prove that compressed models can replace larger systems across every workload. Buyers still need independent testing for accuracy, reliability, energy use and security. But Bonsai 2 makes model size a strategic deployment variable rather than a fixed cost of using capable AI.