Model Details
Jamba 1.5 Mini
Quick answer: Jamba 1.5 Mini is AI21 Labs' August 2024 compact hybrid SSM/Transformer model, scoring 46.1% Arena Hard, 85.7% ARC-C, 75.8% GSM8K, and 42.5% MMLU-Pro. Apache 2.0 licensed with a 256K token context window.
At a Glance
Where Jamba 1.5 Mini leads
- 256K token context window — among the largest for its size class
- Apache 2.0 — fully open for commercial use
- Hybrid SSM/Transformer architecture: efficient long-context inference
- 85.7% ARC-C — strong commonsense reasoning for a mini model
- Compact model with long-context efficiency
Where it lags
- 42.5% MMLU-Pro — below larger models
- 75.8% GSM8K — moderate math
- 46.1% Arena Hard — competitive but below frontier models
- August 2024 release — older than current SOTA
Best for: Applications requiring long-context efficiency (RAG, document analysis, summarisation) in a self-hosted Apache 2.0 compact model.
What Jamba 1.5 Mini Is
Jamba 1.5 Mini is the compact model in AI21 Labs' Jamba 1.5 series, released August 2024 alongside the larger Jamba 1.5 Large. Both use AI21's hybrid SSM (State Space Model) and Transformer architecture — called Mamba-Transformer or Jamba architecture — which enables efficient inference over very long contexts compared to pure-attention models.
The "Mini" designation reflects a smaller parameter count with maintained 256K context, making it well-suited for deployments where both long context and low compute cost are required. Its Apache 2.0 license (matching Jamba 1.5 Large) allows full commercial use including fine-tuning and redistribution.
Specifications
| Field | Value |
|---|---|
| Organization | AI21 Labs |
| License | Apache 2.0 |
| HuggingFace | ai21labs/AI21-Jamba-1.5-Mini |
| Release date | August 22, 2024 |
| Modality | Text only |
| Context window | 256K tokens |
| Architecture | Hybrid SSM/Transformer (Jamba) |
Pricing
Open weights under Apache 2.0 — self-host at no license cost. Also available via AI21 Studio API and major cloud providers.
Public Benchmark Scores
| Benchmark | Score | Source | Date |
|---|---|---|---|
| Arena Hard | 46.1% | Benchgen evaluation | 2024-08 |
| ARC-C | 85.7% | Benchgen evaluation | 2024-08 |
| GSM8K | 75.8% | Benchgen evaluation | 2024-08 |
| MMLU-Pro | 42.5% | Benchgen evaluation | 2024-08 |
Jamba 1.5 Mini vs Alternatives
| Model | Arena Hard | MMLU-Pro | Context | License |
|---|---|---|---|---|
| Jamba 1.5 Mini | 46.1% | 42.5% | 256K | Apache 2.0 |
| Jamba 1.5 Large | 65.4% | 53.5% | 256K | Apache 2.0 |
| Llama 3.1 8B Instruct | — | — | 128K | Llama 3.1 |
| Gemma 3 12B | — | 60.6% | 128K | Gemma ToU |
Jamba 1.5 Mini vs Jamba 1.5 Large: trade 46.1% → 65.4% Arena Hard for a larger model. For pure long-context efficiency in Apache 2.0: Jamba 1.5 Mini is unique at this context/size ratio.
Related Models & Benchmarks
- Compare: Jamba 1.5 Large · Llama 3.1 8B Instruct · Gemma 3 12B
- Benchmarks: Arena Hard · MMLU-Pro · GSM8K
Frequently Asked Questions
What is Jamba 1.5 Mini?
AI21 Labs' August 2024 compact hybrid SSM/Transformer model with 256K context window, Apache 2.0 license, scoring 46.1% Arena Hard and 42.5% MMLU-Pro.What is the Jamba architecture?
Jamba is AI21's hybrid architecture that combines State Space Model (SSM/Mamba) layers with Transformer attention layers. This enables more efficient inference over long contexts than pure-Transformer models, especially at 256K token lengths.Should I use Jamba 1.5 Mini or Jamba 1.5 Large?
Jamba 1.5 Large for better performance (65.4% Arena Hard vs 46.1%). Jamba 1.5 Mini for lower compute requirements where long-context efficiency is primary.Specs from AI21 Labs' Jamba 1.5 Mini release (August 2024) and Benchgen evaluations. Last updated 2026-07-24.