Open-Source Reasoning Model Jamba Reasoning 3B Released

AI21 Labs has released Jamba Reasoning 3B. It is a 3B parameter open-source reasoning model with a SSM-Transformer hybrid architecture and a 256K token context window, capable of processing up to 1M tokens.

The company claims 2-5x efficiency compared to DeepSeek, Google, Llama, and Microsoft, and distributes it under the Apache 2.0 license. It explains that thanks to its lightweight memory structure, it can run on-device on iPhone, Android, Mac, and PC.

The key performance points are as follows.

• Keeps the KV cache 8x smaller than a typical Transformer, lowering memory usage even in long contexts.
• Processes 40 tokens per second on an M3 MacBook Pro with a 32K token context.
• Claims superior performance over on-device models on IFBench, MMLU-Pro, and Humanity's Last Exam.
• Post-training combined RLVR, SFT, DPO, GRPO with the company's own techniques.

AI21 Labs positions this model for local entity extraction from legal and medical documents, manual access for field technicians, personalized file DB-based conversation and writing assistants, and on-device RAG use cases. Later in the article, it emphasizes that SLM routing—which can handle 40-70% of AI tasks at 10-30x lower cost—is advantageous in terms of cost, latency, and privacy.