Contrastive-LM releases an 8B model that scores agent actions rather than writing them

single source· 1 articles · confidence: low · first seen 2026-09-24 05:27 UTC

What this means for you

If you run an agent loop that pays a large model call to choose each step, this is the shape of a cheaper alternative: one small forward pass per candidate action instead of a generation. Nothing to adopt yet — the write-up gives no licence, no weights link, no evaluation date and no Jev numbers.

Contrastive-LM has released CLM-8B, an 8B-parameter model that does not generate text. It scores candidate actions against a state so the loop around it can pick one. It is a frozen Qwen3-8B encoder plus two small projection heads (narrow trainable layers), trained with a contrastive objective that pulls matching state-action pairs together. Zero-shot, it runs up to 9× faster than TypeSafe's Jev, per the release. With fine-tuned heads used as a verifier, it reports 81.6% on held-out DeepSWE tasks and 87.6% on held-out Terminal-Bench 2.1. No evaluation dates or harness details are given, and this rests on a single write-up.

Key facts

  • ·CLM-8B is an 8B-parameter model that scores candidate actions against a state instead of generating text. source
  • ·It is built by adding two small projection heads to a frozen Qwen3-8B encoder, trained with a contrastive InfoNCE objective. source
  • ·Zero-shot, the release says it runs up to 9× faster than TypeSafe's Jev. source
  • ·With fine-tuned heads used as a verifier it reports 81.6% on held-out DeepSWE tasks. source
  • ·It reports 87.6% on held-out Terminal-Bench 2.1 tasks. source
  • ·No evaluation date or harness is given for either reported score, and no benchmark figures are given for Jev. source

What the sources say

  • MarkTechPost — Release write-up listing the model's construction, its speed comparison with Jev and two held-out accuracy figures.

Sources

The original reporting. Follow these — they did the work.

← the wire