Barak Lenz

Barak Lenz is an AI researcher and technology leader who coauthored the Jamba hybrid language-model papers. AI21 identified him as its chief technology officer in January and March 2026. His public technical argument emphasizes how systems validate answers, choose execution paths, and manage time and cost—not simply how models rank on average.[4][5][2][3]

In the January 15, 2026 YAAP episode “Why AI Leaderboards Miss the Point”, Lenz discussed four gaps between model capabilities and production systems with AI21 developer advocate Yuval Belfer: validation, contextualization, latency, and decomposition. Belfer’s subsequent account of their conversation quotes Lenz’s objection to compressing the problem into a slogan: “I don’t like one-liners. They tend to throw away the specific details that actually explain what’s going on.”[2][3]

The validation problem is that generating a correct answer among several attempts does not ensure a system can identify it. Contextualization adds another choice: which model or approach suits this particular input? As Belfer quotes Lenz, “This gap is the difference between what’s best on average and what’s best for a specific input.” Belfer attributes to him a portfolio of actions, governed by an action model that estimates the value and cost of possible paths before taking them. Easy inputs could go to cheaper models, with additional spending reserved for harder work.[3]

Lenz also distinguishes dollar cost from latency. In Belfer’s account, optimizing completion time calls for parallel attempts and early termination of remaining work once an answer passes validation. Decomposition extends those choices through a workflow: intermediate steps become opportunities to branch, stop, or route a subtask elsewhere, rather than committing to one sequential trajectory. Lenz’s proposed “AI Operating System” would manage these execution decisions and resources.[3]

Before these 2026 discussions of execution strategy, Lenz’s collaborative research addressed language-model architecture and training. The 2024 Jamba papers provide context for his work on models with explicit memory and inference constraints.[4][5][2]

Lenz coauthored “Jamba: A Hybrid Transformer-Mamba Language Model”, first submitted on March 28, 2024, with Opher Lieber, Hofit Bata, and 19 other collaborators. Jamba interleaves Transformer and Mamba layers and adds mixture-of-experts capacity in some layers while keeping active parameter usage manageable. The authors describe an architecture configurable for different resources and objectives; they reported a configuration that fit on a single 80GB GPU and strong results at context lengths up to 256K tokens. They also made the implementation’s weights publicly available under a permissive license.[4]

The Jamba-1.5 paper, submitted August 22, 2024, lists Lenz among a 60-person team. It introduced instruction-tuned versions of the hybrid architecture and ExpertsInt8 quantization. The team reported that this technique allowed Jamba-1.5-Large to process 256K-token contexts on eight 80GB GPUs without loss of quality. The model weights were released under the Jamba Open Model License, while ExpertsInt8 was released as open source.[5]

Lenz also appeared on Latent Space in “Building Jamba 3B: the tiny Hybrid Transformer State Space Reasoning Model”, which introduced him as AI21’s CTO.[1]

Earlier, Lenz coauthored “SenseBERT: Driving Some Sense into BERT”, published at ACL in July 2020, with Yoav Levine, Or Dagan, Ori Ram, Dan Padnos, Or Sharir, Shai Shalev-Shwartz, Amnon Shashua, and Yoav Shoham. Their model was pretrained to predict both masked words and their WordNet supersenses, adding weak supervision at the word-sense level without human annotation. The team reported improved lexical understanding on word-sense disambiguation and Word in Context tasks.[6]

Lenz’s language-model publications also include the Jurassic-1 technical white paper, coauthored with Opher Lieber, Or Sharir, and Yoav Shoham. The paper identifies his affiliation as AI21 Labs.[7]