CLM-8B: contrastive System One model claims 9x faster inference, agentic SOTA
ChengleiSi · x · 2026-09-24
A new Contrastive Language Model (CLM) uses a contrastive objective connecting states and actions.
- CLM-8B, pre-trained on internet-scale data, delivers up to 9x faster inference than Jev with comparable performance on computer-use, gaming, and tool-calling tasks
- With lightweight fine-tuning it sets SOTA on agentic coding benchmarks: DeepSWE (81.6%) and Terminal-Bench 2.1 (87.6%); Jev fails as an effective verifier for these long-horizon tasks
- Training/serving infra disaggregates states and actions so embeddings can be cached independently, cutting latency when state evolves but the action set is fixed
More from Models
- One-shot with Opus 5.5 recreates a playable 1990s record store in 3D — petergyang · 2026-09-24
- Claude 5.6 at medium effort correctly merges, OCRs and vectorizes whiteboard photos — iskander · 2026-09-24
- Four whiteboard photos in, Astra merges, OCRs and vectorizes them correctly in one shot — iskander · 2026-09-24
- Developer says Claude usage shifted from Fable limits to mostly Opus limits this week — rudrank · 2026-09-24
- Users call out OpenAI: Sol looks like rebranded Terra with $2 cheaper output but worse usage limits — RexDouglass · 2026-09-24
- That Opus 5.5 Demoscene Demo Used Under 3% of a Weekly Usage Quota — gandamu_ml · 2026-09-24