Jamba Reasoning 3B
Jamba Reasoning 3B uses 28 layers: 26 Mamba layers and two attention layers. Its compact architecture targets reasoning without the weight size of Jamba Mini or Large.
How reasoning was added
AI21 used about half a trillion mathematics and code tokens for mid-training. Later stages included reasoning distillation, preference optimization and online reinforcement learning with verifiable rewards. The training tasks included code, mathematics, structured output and extraction.AI21-Jamba-Reasoning-3B official model card
The GGUF release contains the same model
AI21 publishes a separate GGUF repository for local inference. It is a distribution format of Reasoning 3B, not a second model generation.AI21-Jamba-Reasoning-3B-GGUF official model card
Practical guidance. A quantized file can reduce memory use, but it can also change output quality. Compare the chosen format with the original weights on multi-step problems and long inputs.
Hybrid inference needs suitable kernels
The model card recommends vLLM 0.11 or later and a float32 Mamba cache. These settings matter because this architecture stores recurrent state as well as attention data.AI21-Jamba-Reasoning-3B official model card
Practical guidance. Reasoning text is still generated text. Check final answers and code execution separately instead of treating a longer explanation as evidence of correctness.
Specifications & access
| Detail | Value and conditions |
|---|---|
| Launch date | |
| Access | Downloadable |
| Model identifiers | AI21-Jamba-Reasoning-3Bai21labs/AI21-Jamba-Reasoning-3BAI21-Jamba-Reasoning-3B-GGUF |
| Context window | 262,144 tokens Maximum position count in the published weight configuration; deployed limits depend on the serving setup. AI21-Jamba-Reasoning-3B configuration |
| Parameters | 3 billionAI21-Jamba-Reasoning-3B official model card |
| Architecture | 26 Mamba layers; 2 attention layers; 64K vocabularyAI21-Jamba-Reasoning-3B official model card |
| Weight license | Apache 2.0, as stated in the model cardAI21-Jamba-Reasoning-3B official model card |
Related models
Compact dense language model for private document processing and instruction-based tasks.
Launch date Downloadable- Jamba2 Mini
AI21 Labs · Jamba2
Hybrid language model with 12 billion active parameters for document questions, extraction and structured text.
Launch date API · Downloadable