AI21 Jamba 1.5 Mini

Acerca de AI21 Labs

Fundado en 2017 · Tel Aviv, Israel

AI21 Labs is an Israeli LLM pioneer founded in 2017 by Yoav Shoham (Stanford emeritus), Ori Goshen and Amnon Shashua (Mobileye founder). Long active in commercial LLMs (Jurassic-1, Jurassic-2), AI21 pioneered the hybrid State-Space + Transformer Jamba architecture in March 2024. Jamba 1.5 Mini is the small-and-fast variant of the August 2024 Jamba 1.5 release, optimised for long-context inference at production cost points and tractable on a single H100 80GB. AI21 has raised over $336M from investors including Google, Nvidia, Walden Catalyst and Pitango.

Arquitectura

Hybrid Mamba-Transformer Mixture-of-Experts

Jamba 1.5 Mini uses the same hybrid SSM+Transformer+MoE recipe as Jamba 1.5 Large at smaller scale. Across 32 blocks each block alternates Mamba (selective state-space) and self-attention layers in a 7:1 ratio. MLPs are MoE with 16 experts and top-2 routing, giving 12B active parameters out of 52B total. The 12B active count means inference cost is competitive with dense 12B models, while Mamba layers provide constant-memory long-context scaling. The model uses a 64,000-token BPE tokeniser and supports the same nine languages as Jamba 1.5 Large (English, Spanish, French, Portuguese, Italian, Dutch, German, Arabic, Hebrew). Released August 2024 under the Jamba Open Model License with hosted access via AI21 Studio, AWS Bedrock, Azure AI Studio and Snowflake Cortex.

Parámetros

52B total, 12B active per token (16 experts, top-2 routing)

Contexto

256.000 tokens

Capacidades

• Hybrid Mamba+Transformer+MoE architecture at compact scale
• 52B total / 12B active parameters
• 256K effective context
• Fits on a single H100 80GB at FP16 or A100 at INT8
• Native function calling and JSON-mode output
• Multilingual (9 languages)
• Open weights under Jamba Open Model License
• Best for: cheap long-context summarisation, RAG with large retrieval windows, single-GPU enterprise pilots.

Entrenamiento y licencia

Same data mixture and methodology as Jamba 1.5 Large: trillions of tokens of web, code, math, books and multilingual sources, knowledge cutoff March 2024, followed by supervised fine-tuning and preference optimisation.

Licencia: Jamba Open Model License. Permits research and commercial use with attribution and AUP compliance. Hosted access via AI21 Studio, AWS Bedrock, Azure AI Studio and Snowflake Cortex.

Pruebas de seguridad: AI21 publishes a model card with safety and limitation disclosures. Positioning is enterprise-focused (grounded, structured outputs) rather than consumer-chat safety tuning.