Model Details
Jamba 1.5 Large
Quick answer: Jamba 1.5 Large is AI21 Labs' August 2024 hybrid SSM-Transformer model with a 256K context window, scoring 65.4% Arena Hard, 87.0% GSM8K, and 53.5% MMLU-Pro. It is unique for combining Mamba SSM layers with Transformer attention — offering linear-time long-context processing with Apache 2.0 licensing.
At a Glance
Where Jamba 1.5 Large leads
- 256K context window — exceptionally long at launch
- Hybrid SSM-Transformer: linear time complexity for long contexts
- Apache 2.0 — fully permissive commercial use
- 87.0% GSM8K — strong math at this scale
- Efficient long-context inference (vs quadratic cost of pure Transformers)
Where it lags
- 65.4% Arena Hard — below frontier transformer-only models
- 53.5% MMLU-Pro — moderate academic knowledge
- Architecture complexity: SSM-Transformer requires specific inference support
- March 2024 knowledge cutoff
Best for: Long-document processing requiring 256K context with permissive licensing; research into hybrid SSM-Transformer architectures; teams needing Apache 2.0 at 256K context.
What Jamba 1.5 Large Is
Jamba 1.5 Large (released August 22, 2024) is AI21 Labs' flagship open-weight model combining Mamba SSM (State Space Model) layers with standard Transformer attention layers. This hybrid architecture allows efficient processing of very long contexts — linear rather than quadratic time complexity — making it particularly suitable for 256K-token document analysis.
The SSM-Transformer hybrid is architecturally significant: Mamba layers replace many attention layers, reducing KV cache size and making long-context inference more memory-efficient. This allows the model to handle 256K-token inputs (roughly 190 pages) with less GPU memory than a pure Transformer at the same context length.
At 65.4% Arena Hard and 53.5% MMLU-Pro, general instruction following and knowledge benchmark scores are below comparable dense transformers (Llama 3.3 70B: 68.9% MMLU-Pro). The tradeoff is architectural: Jamba delivers better long-context efficiency, but pure Transformer models generally score higher on standard benchmarks.
Specifications
| Field | Value |
|---|---|
| Organization | AI21 Labs |
| Context window | 256,000 tokens |
| License | Apache 2.0 |
| HuggingFace | ai21labs/AI21-Jamba-1.5-Large |
| Release date | August 22, 2024 |
| Knowledge cutoff | March 2024 |
| Architecture | Hybrid SSM-Transformer (Mamba + Attention) |
| Modality | Text only |
Pricing
Open weights under Apache 2.0 — free to self-host. Available via AI21 Labs API.
Context Window
Jamba 1.5 Large supports a 256,000-token context window — approximately 190 pages of text. This was among the longest context windows at launch for an open-weight model.
Public Benchmark Scores
| Benchmark | Score | Source | Date |
|---|---|---|---|
| Arena Hard | 65.4% | Benchgen evaluation | 2025-07 |
| GSM8K | 87.0% | Benchgen evaluation | 2025-07 |
| MMLU-Pro | 53.5% | Benchgen evaluation | 2025-07 |
Jamba 1.5 Large vs Alternatives
| Model | MMLU-Pro | Context | License | Architecture |
|---|---|---|---|---|
| Jamba 1.5 Large | 53.5% | 256K | Apache 2.0 | SSM+Transformer |
| Llama 3.3 70B Instruct | 68.9% | 128K | Llama 3.3 | Transformer |
| Llama 3.1 70B Instruct | — | 128K | Llama 3.1 | Transformer |
Jamba 1.5 Large vs Llama 3.3 70B: lower MMLU-Pro (53.5% vs 68.9%) but 2× context window (256K vs 128K) under Apache 2.0. Choose Jamba for long-document tasks; choose Llama 3.3 for general-purpose quality.
Related Models & Benchmarks
- Compare: Llama 3.3 70B Instruct · Qwen2.5 72B Instruct
- Benchmarks: Arena Hard · GSM8K · MMLU-Pro
Frequently Asked Questions
What is Jamba 1.5 Large?
Jamba 1.5 Large is AI21 Labs' August 2024 hybrid SSM-Transformer model scoring 65.4% Arena Hard, 87.0% GSM8K, and 53.5% MMLU-Pro with a 256K context window under Apache 2.0.What makes Jamba 1.5 Large different?
Jamba uses a hybrid SSM-Transformer architecture combining Mamba SSM layers with Transformer attention. This enables linear-time long-context processing, making 256K-token inference more memory-efficient than pure Transformers.Is Jamba 1.5 Large open source?
Yes — Apache 2.0. Free for commercial use without restrictions.Specs from AI21 Labs' official Jamba 1.5 release (August 2024) and Benchgen evaluations. Last updated 2026-07-24.