AI21 Labs Releases Jamba Reasoning 3B for On-Device AI Workloads

AI21 Labs has launched Jamba Reasoning 3B, a compact open-source reasoning model designed for on-device AI workloads, announced in a company blog post. The 3-billion-parameter model is built on a hybrid SSM-Transformer architecture, offering efficiency gains of two to five times compared to similar models from DeepSeek, Google, Meta, and Microsoft.

Jamba Reasoning 3B supports a context window of 256,000 tokens, with the ability to handle up to 1 million. According to the company, the model can generate 40 tokens per second on an M3 MacBook Pro at a 32,000-token context length. Licensed under Apache 2.0, it is available for download and local inference via Hugging Face, Kaggle, LM Studio, and llama.cpp.

The model’s architecture reduces memory usage through an optimized KV cache, making it suitable for enterprise and personal applications that require secure, offline operation. AI21 Labs highlighted that small language models like Jamba Reasoning 3B can handle 40–70% of enterprise AI tasks at up to 30 times lower cost compared to large cloud-based models.

In addition to efficiency, Jamba Reasoning 3B demonstrated leading performance on benchmarks such as IFBench, MMLU-Pro, and Humanity’s Last Exam. The company said it plans to continue expanding the Jamba model family, focusing on enabling cost-effective, decentralized AI computation across devices.