Stanford and Nvidia's CLM-8B model runs up to 9x faster than Jev

by@VentureBeat

Summary

Researchers from Stanford and Nvidia have unveiled CLM-8B, a new Contrastive Language Model (CLM) designed to enhance decision-making speed and efficiency in AI applications, caching agent actions for faster retrieval. This innovative model processes agent actions by creating state and action embeddings, allowing it to select the most relevant action without generating token sequences, and tests have shown it runs up to nine times faster than TypeSafe's Jev in various tasks. Its dual-encoder architecture is particularly effective for systems requiring predefined action sets, marking a significant advancement in how bounded decision tasks are approached. The CLM-8B weights, code, and tools are now available under the Apache 2.0 license, enabling immediate integration for developers.

Analysis

Jev: Jev is TypeSafe’s model for rapid, calibrated decisions in agent workflows, trained with Reinforcement Learning for Calibrated Decisions. It serves as the primary baseline in evaluations of CLM-8B across gaming, tool calling, and computer-use tasks. Nvidia: Nvidia partnered with Stanford researchers to create and release the open CLM-8B model. The collaboration focuses on advancing efficient inference patterns for agentic workflows through contrastive representations of states and actions. Qwen3-8B: Qwen3-8B provides the frozen backbone for the released CLM-8B model, with separate projection heads added for state and action embeddings. This architecture supports caching of action representations to accelerate repeated decisions without retraining the full model. Stanford: Stanford University researchers collaborated with Nvidia on the development of Contrastive Language Models. The team released CLM-8B, an 8-billion-parameter model designed for efficient decision-making in AI agents. Project lead Jacky Kwok highlighted advantages of the contrastive training approach for domain-specific tasks. TypeSafe: TypeSafe introduced the Jev model earlier this month as part of its push for “System One” models focused on fast, structured decisions. The company positions CLM as a complementary contrastive approach within this emerging category of decision systems. Jacky Kwok: Jacky Kwok serves as the project lead from Stanford on the CLM initiative. He has discussed how the contrastive objective outperforms standard supervised fine-tuning for decision-making and enables quick adaptation to downstream tasks via frozen backbones. Release: CLM-8B weights, code, API compatibility, and fine-tuning tools have been made available under the Apache 2.0 license for immediate developer use. Innovation: CLM reframes bounded decision tasks as a contrastive matching problem between state and action embeddings rather than token generation. Architecture: The dual-encoder design allows independent caching of state and action representations, suiting applications with reusable or long candidate action sets.

Categories

machine_learningai_agentsaitech
View Original Tweet