Open-Weight CLM Released as Jev Alternative, Agent Decision-Making 4x Faster

The open-weight model CLM (Contrastive Language Models), designed to replace TypeSafe AI's closed-source decision-making model Jev, has been released. CLM is based on Qwen3-8B and provides the same API and functional interface as Jev.

Performance and Architectural Improvements

CLM adopts an architecture that separates the state head and action head, allowing it to cache repetitive action embeddings. This results in 4x to 13x faster latency compared to Jev in interactive browser agent and game environment benchmarks. Additionally, its small parameter size of approximately 75MB enables fine-tuning to match users' agent trajectories.

Benchmark Results

The fine-tuned CLM achieved state-of-the-art performance in coding benchmark verifier tasks.

• Terminal-Bench 2.1: 87.6%
• DeepSWE: 81.6%

In contrast, the zero-shot Jev scored approximately 71% on DeepSWE, showing relatively lower performance.

Limitations and Differences

While CLM supports all of Jev's core primitives (Choice, Noul, Score), there are some limitations.

• Generalization Ability: Jev demonstrates higher accuracy in zero-shot open-domain tasks (e.g., Jev 99.2% vs CLM-8B 95.2% on the Berkeley Function Calling Leaderboard v4).
• Context Length: Jev natively supports a 64K token context, whereas CLM-8B has been validated and calibrated for 2K–8K contexts.
• Probability Normalization: CLM's probabilities are relative to the candidate set provided at request time, while Jev provides internally calibrated scores against absolute criteria.