Atlas

Models

← All models

Jamba 1.5 Large

AI21 Labs · 2024-08-22 · 398.6B parameters

Jamba 1.5 Large is AI21 Labs' instruction-tuned hybrid Transformer-Mamba mixture-of-experts model, released on August 22, 2024, with nine eight-layer blocks, a 1:7 attention-to-Mamba ratio, 16 experts, and top-2 routing. AI21 reports approximately 398B total and 94B active parameters and a 256K-token effective context window; the model supports function calling, structured JSON output, grounded generation, and nine languages.

Benchmark scores

BenchmarkScore
Artificial Analysis Intelligence Index5.1
Capability117.8
CorpFin v239.4
CorpFin v2 - Exact Pages45.5
CorpFin v2 - Max Fitting Context36.4
CorpFin v2 - Shared Max Context36.5
GPQA Diamond42.7
Humanity's Last Exam (Text-Only)4.0
LegalBench74.2
LegalBench - Conclusion Tasks79.7
LegalBench - Interpretation Tasks74.9
LegalBench - Issue Tasks72.8
LegalBench - Rhetoric Tasks73.0
LegalBench - Rule Tasks70.4
LiveCodeBench (Artificial Analysis)14.3
MATH-500 (Artificial Analysis source)60.6
MedQA Demographic Bias (Vals)68.1
MedQA Demographic Bias (Vals) - Asian67.7
MedQA Demographic Bias (Vals) - Black67.6
MedQA Demographic Bias (Vals) - Hispanic68.2
MedQA Demographic Bias (Vals) - Indigenous67.2
MedQA Demographic Bias (Vals) - Unbiased69.7
MedQA Demographic Bias (Vals) - White68.5
MMLU78.2
MMLU Pro57.2
SciCode (Artificial Analysis)16.3
TaxEval v258.2
TaxEval v2 - Correctness46.8
TaxEval v2 - Stepwise Reasoning69.6
Loading Atlas data…