AI21 Labs logo

Jamba 1.5 Large

AI21 Labs

Jamba 1.5 Large is AI21's flagship language model based on the Jamba architecture, which combines Transformer and Mamba layers for efficient long-context processing. It delivers strong performance in reasoning, knowledge, and conversation while supporting extended context windows with high throughput.

Key Specifications

Parameters
398.0B
Context
256.0K
Release Date
August 22, 2024
Average Score
65.5%

Timeline

Key dates in the model's history
Announcement
August 22, 2024
Last Update
July 19, 2025
Today
March 25, 2026

Technical Specifications

Parameters
398.0B
Training Tokens
-
Knowledge Cutoff
March 5, 2024
Family
-
Capabilities
MultimodalZeroEval

Pricing & Availability

Input (per 1M tokens)
$2.00
Output (per 1M tokens)
$8.00
Max Input Tokens
256.0K
Max Output Tokens
256.0K
Supported Features
Function CallingStructured OutputCode ExecutionWeb SearchBatch InferenceFine-tuning

Benchmark Results

Model performance metrics across various tests and benchmarks

General Knowledge

Tests on general knowledge and understanding
MMLU
Accuracy : 1. Accuracy : 2. Accuracy : accuracyaccuracyaccuracyaccuracySelf-reported
81.2%
TruthfulQA
Accuracy AI: ChatGPT is a language model that can solve questions by processing patterns in language.Self-reported
58.3%

Mathematics

Mathematical problems and computations
GSM8k
Accuracy AI ## Score Accuracy (AIME, GPQA), accuracy ### * AIME * GPQA * * ### * ****: * ****:Self-reported
87.0%

Reasoning

Logical reasoning and analysis
GPQA
Accuracy AI: 0Self-reported
36.9%

Other Tests

Specialized benchmarks
ARC-C
Accuracy AI: *no output*Self-reported
93.0%
Arena Hard
Accuracy AI: ChatGPT was asked to solve 100 questions from MMLU on tasks including elementary mathematics, US history, computer science, and law. The model achieved an accuracy of 86.7%. This accuracy is compared against human expert performance (89.8%) and previous state-of-the-art models (Gemini Ultra: 83.7%, Claude 2: 78.5%). Results breakdown: - Elementary mathematics: 92.3% (vs human: 95.1%) - US history: 84.5% (vs human: 87.2%) - Computer science: 88.9% (vs human: 91.4%) - Law: 81.1% (vs human: 85.5%) The model performs consistently across domains, with strongest results in mathematical reasoning tasks. Error analysis shows that mistakes primarily occurred on questions requiring specialized knowledge rather than general reasoning capabilities.Self-reported
65.4%
MMLU-Pro
Accuracy : - : - : MATH, GSM8KSelf-reported
53.5%
Wild Bench
Accuracy AI: Accuracy ("26.83" "26 + 0.83"). accuracy evaluation accuracySelf-reported
48.5%

License & Metadata

License
jamba_open_model_license
Announcement Date
August 22, 2024
Last Updated
July 19, 2025

Similar Models

All Models

Recommendations are based on similarity of characteristics: developer organization, multimodality, parameter size, and benchmark performance. Choose a model to compare or go to the full catalog to browse all available AI models.