AI21

System automatically learns the unique environment and inputs, adapting as things change

AI21 offers Jamba Models described as Efficient LLMs for long-context processing.

AI21 offers Maestro, an optimization framework for real-world AI agents.

AI21 achieved a state-of-the-art 60.9% issue resolve rate on the SWE-rebench benchmark across 123 issues between December and March.

AI21's Maestro optimization approach has been applied across multiple agent benchmarks, including SWE-bench Verified, BrowseComp-Plus, and Deep Research Bench.

AI21 employs Algorithm Developers, including Eli Lepkifker and Oded Avraham.

Jamba Models are described as efficient LLMs for long-context processing

AI21's blog covers advancements in natural language processing and machine learning

AI21 maintains a research lab called 'Inside The Lab'

AI21 offers a newsletter providing enterprise AI news, product developments, customer success stories, and GenAI updates

Pioneers enterprise AI systems and foundation models with a mission of trustworthy AI that powers superproductivity

Prof. Yoav Shoham is a professor emeritus of computer science at Stanford University and Google's former Principal Scientist

Ori Goshen has over 15 years of experience in technology and product leadership roles, with prior achievements including co-founding network analytics company Crowdx and spearheading VoIP development

Maestro is an optimization framework for real-world AI agents

Jamba Models are efficient LLMs for long-context processing

Maestro is an optimization framework for real-world AI agents

Jamba Models are efficient LLMs for long-context processing

Achieved state-of-the-art (SOTA) performance with 95.18% accuracy on the BrowseComp-Plus benchmark using AI21 Maestro

Maestro is used to optimize deep research agents, balancing accuracy, cost, and latency

BrowseComp-Plus tests retrieval precision and synthesis across deep corpus search tasks; Deep Research Bench 1 tests long-form report generation and quality

Amnon Morag is VP Product at AI21

Jamba Models are efficient LLMs for long-context processing

Maestro is a general-purpose agentic framework that automatically scales compute and optimizes orchestration

Maestro uses structured plans, automatic horizontal scaling, and decision-theoretic optimization techniques to allocate test-time compute

Maestro employs structured Test-Time Compute mechanisms that explicitly and adaptively allocate resources during execution

AI21 operates a research division called 'The Lab'

AI21 Maestro executes multiple trajectories concurrently, exploring different approaches and selecting the best outcome.

Answering complex questions based on an internal knowledge base using parallel reasoning agents.

Uses MCP (Model Context Protocol) to handle tool calls such as running shell commands, compiling, running tests, and inspecting outputs.

Parallel agents work well for reasoning but break down once they begin acting, because multiple branches editing the same files cause conflicts and corruption.

Jamba Models — Efficient LLMs for long-context processing.

Article authored by Eran Gat, System Lead, at AI21.