CLM-v0.1-8B

CLM-v0.1-8B is the first contrastive language model from the Contrastive-LM research group, announced on 23 September 2026. It does not generate text. Scoring is done in one pass, so an AI agent can rank candidate moves or verify a planned move without producing any output text.

Features

• Action scoring: scores how well each candidate action fits a given situation.
• Step checking: ranks possible next steps, or checks one proposed step, in a single pass.
• Frozen encoder: the core is a frozen Qwen3-8B encoder, paired with two small trained heads, one reading situations and one reading actions.
• Verifier results: as a fine-tuned verifier on held-out subsets, the authors report 81.6 per cent for DeepSWE and 87.6 for Terminal-Bench 2.1.
• Latency: up to nine times lower than Jev when used without training, per the authors.

Customers and markets

Several open alternatives to Jev appeared within about two weeks, which suggests the approach is not tied to one supplier.

Key facts

• Licence: Apache 2.0, which allows commercial use.
• Trained part: about 20 million parameters, around 75 MB in addition to the base model, according to MarkTechPost.
• Tool-calling success: 95.2 per cent, against 99.2 for Jev, per the authors.