# Jamba 1.5 Large ![Organization](https://img.shields.io/badge/Org-AI21-6366f1) ![Context](https://img.shields.io/badge/Context-256K%20tokens-1f6feb) ![License](https://img.shields.io/badge/License-Apache%202.0-2da44e) ![Weights](https://img.shields.io/badge/Weights-Open-2da44e) ![Released](https://img.shields.io/badge/Released-Aug%202024-555) ![Architecture](https://img.shields.io/badge/Architecture-SSM%2BTransformer-8250df) > **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](https://huggingface.co/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](/benchmarks/arena-hard-v2) | 65.4% | Benchgen evaluation | 2025-07 | | [GSM8K](/benchmarks/gsm8k) | 87.0% | Benchgen evaluation | 2025-07 | | [MMLU-Pro](/benchmarks/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](/models/meta/llama-3-3-70b-instruct) | 68.9% | 128K | Llama 3.3 | Transformer | | [Llama 3.1 70B Instruct](/models/meta/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](/models/meta/llama-3-3-70b-instruct) · [Qwen2.5 72B Instruct](/models/qwen/qwen2-5-72b-instruct) - Benchmarks: [Arena Hard](/benchmarks/arena-hard-v2) · [GSM8K](/benchmarks/gsm8k) · [MMLU-Pro](/benchmarks/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.*

Jamba 1.5 Large — Training jobs

Training jobs