AI BriefingAI21 Labs news
The latest AI and developer news from AI21 Labs, with short summaries.
Visit website →Debugging a Mamba Bug in vLLM
The cause of rare gibberish was traced to vLLM's Mamba scheduler and cache.
Query Dependent Chunking
Noting that the optimal chunk size differs for each query, combining chunks of multiple sizes improved RAG retrieval performance by up to 37%.
Dynamic Data Snoozing
By temporarily putting overly easy examples to sleep in GRPO, this approach reduces wasted computation and achieves up to 3X efficiency.
RTL PDF Document Parsing
Word Shape Encoding significantly boosts Hebrew and Arabic PDF parsing quality.
Being Boring Isn't Easy
AI21 has proposed Boring Agents, which prioritize predictability and auditability.
Introducing Jamba2
AI21 has open-sourced Jamba2, splitting it into two models: 3B and Mini.
Scaling agentic evaluation for SWE-bench
Surviving 200,000 evaluations means even agentic benchmarks need multi-tenant infrastructure.
MCP for Stateful Agent Workspaces
Adding a workspace layer to MCP safely parallelizes agents that mutate state.
Test-Time Compute for SWE-bench
AI21 Maestro boosts SWE-bench performance through structured planning and parallel execution.
Enterprise AI After the Hype
In 2025, the deciding factor in enterprise AI was system design, not model performance.
Collaboration between AI21 and Together AI
AI21 and Together AI announced a collaboration for enterprise open-source AI operations.
AI21 Labs Named an Emerging Visionary in Two Gartner EMQ Categories
AI21 Labs has been named an Emerging Visionary in two Gartner EMQ categories.
Improving Enterprise Accuracy with Structured RAG
AI21 announced that Structured RAG improves enterprise query accuracy by up to 60%.
Jamba 3B vs Qwen3 4B Comparison
Jamba Reasoning 3B was much faster than Qwen3 4B on a 60K-token QA task.
AI21 Maestro, a Reliable AI Agent
AI21 Maestro helps build reliable AI agents by automating everything from data connection to verification.
5 Questions, 5 Answers on Medical AI
Medical AI expands prediction and decision-making, but it cannot replace empathy.
Open-Source Reasoning Model Jamba Reasoning 3B Released
AI21 has released Jamba Reasoning 3B, a 3B-class on-device open-source reasoning model.
AI in Financial Services
AI adoption is spreading across financial services, but trust and governance remain the key to moving into production.
Enterprise Knowledge Agents
Enterprises need execution-capable knowledge agents, not simple RAG chatbots.
The Rise of Hybrid LLMs
Since Mamba, hybrid LLMs have been replacing the limits of Transformers.