Contrastive-LM released CLM-v0.1-8B, a System One model. It uses a frozen Qwen3-8B encoder with two projection heads. Training used 60M Q&A pairs, 30M hard negatives and 1M agentic trajectories. Zero-shot performance matches Jev on computer-use, gaming and tool-calling tasks. It achieves 81.6% on DeepSWE and 87.6% on Terminal-Bench 2.1 when fine-tuned. Latency is up to 9× lower than Jev. Fine-tuning is cheap because only the heads are trained. State and action caching makes it 13× faster than Jev with ~1k candidates. A 35B multimodal version is coming in early October. The model is Apache 2.0 licensed. It only scores candidates and does not generate text. Verifier results require fine-tuning. The encoder is locked to Qwen3-8B embeddings.
What changed
No prior public release of a Contrastive Language Model existed.
What it unlocks
Scoring and ranking candidates for decision-making with lower latency and fine-tunable heads.
- 60M Nemotron Q&A pairs pre-trained
- 30M synthetic hard negatives mid-trained
- 1M agentic trajectories post-trained
- 81.6% on DeepSWE
What you need to act on it
- Qwen3-8B embeddings
- fine-tuning for SOTA verifier performance
Sources