CLM v0.1-8B
Contrastive Language Models score a state against a set of candidate actions with a contrastive objective. Two projection heads on frozen Qwen3-8B embeddings, trained with InfoNCE; clm-serve maps Choice, Score and Noul onto candidate ranking.
The repository holds only the two heads (about 75 MB); the frozen Qwen3-8B encoder is fetched separately and needs a vLLM pooling server on Linux with an NVIDIA GPU. Training used roughly 60M question–answer pairs, 30M hard negatives and 1M agent trajectories. The authors claim parity with Jev on computer-use, gaming and tool-calling, up to 9× lower latency, and, after fine-tuning the heads, 87.6% on Terminal-Bench 2.1 and 81.6% on DeepSWE; none of these are independently reproduced. The server exposes /v1/systemone and /v1/rank; confidence is the top probability minus the mean of the rest, and states are rendered as prose. A multimodal CLM-35B is announced for October.
What it decides
- choice — picks one option from a set
- score — places the input on an ordered scale
- noul — answers a yes/no question with one calibrated probability
- rank — orders a set of items
- classify — assigns a category from a fixed taxonomy
- route — sends the input to one of several destinations
At a glance
| Parameters | 8B |
| Base model | qwen/qwen3-8b |
| Maker | Contrastive-LM |
| Released | 2026-09-23 |
| License | apache-2.0 |
| Reported latency | authors: up to 9× lower than Jev; ~28 ms for a new state on an RTX 4090 |
Get the weights
pip install systemonemodels
systemone pull contrastive-lm/clm
The files are served from the maker's Hugging Face repository, Contrastive-LM/CLM-v0.1-8B, and verified against the checksums recorded here.
Read more
This page was opened by System One for Contrastive-LM, who can claim the organisation and take it over at any time.