Atlas

Models

← All models

Jamba Reasoning 3B

AI21 Labs · 2025-10-08 · 3.2B parameters

Jamba Reasoning 3B is AI21 Labs' compact open-weight reasoning model, publicly released on October 8, 2025. Its 3,197,109,632-parameter checkpoint uses a 28-layer hybrid Transformer–Mamba architecture with 26 Mamba layers and two MQA attention layers, and is configured for a 262,144-token context; AI21 separately reports that the model can handle up to 1 million tokens. Intended for edge and on-device inference, its published training stages include large-scale pretraining, approximately 0.5 trillion tokens of math-and-code mid-training, SFT, DPO, and RLVR for reasoning, tool use, instruction following, code generation, mathematics, structured output, and information extraction.

Benchmark scores

BenchmarkScore
AA-LCR7.0
AA-Omniscience Index-73.7
AIME 202510.7
Artificial Analysis Intelligence Index4.1
Artificial Analysis Omniscience Accuracy6.6
Artificial Analysis Omniscience Hallucination Rate86.0
Artificial Analysis Openness Index38.9
Capability117.3
CritPt0.0
GPQA Diamond33.3
Humanity's Last Exam (Text-Only)4.6
IFBench (Artificial Analysis)52.4
LiveCodeBench (Artificial Analysis)21.0
MMLU Pro57.7
SciCode (Artificial Analysis)5.9
Terminal-Bench Hard0.8
τ²-bench (Telecom)15.8
Loading Atlas data…