Oriven Labs
Oriven Labs is an independent AI lab building open models for coding and agents.
DAMS
DAMS (Dynamic Associative Memory System) is our open model family for software engineering and long-horizon agent work.
DAMS 27B (first release)
- Full-parameter SFT and RL on Qwen3.8-27B (Gated DeltaNet hybrid). Every weight trained, no LoRA.
- RL rewards come only from real repository tests passing.
- Associative memory layers added on top of the base architecture.
- Open weights, Apache 2.0.
| Benchmark (pass@1) |
DAMS 27B |
Qwen3.8-27B (base) |
| LiveCodeBench v6 |
88.2 |
84.8 |
| GPQA Diamond (avg@8) |
90.0 |
89.2* |
| SWE-bench Verified |
85.2 |
86.6 |
| DeepSWE v1.1 |
39.8 |
44.2 |
LiveCodeBench, SWE-bench Verified (mini-SWE-agent, 250 steps), and DeepSWE (Claude Code, 3 h per task) were run by Oriven Labs in identical harnesses for both models. *GPQA base score as published by Qwen. We report wins and losses as measured.
Principles
- Open. Weights and model cards are released under permissive licenses.
- Honest evals. Decontaminated training data. pass@1, pass@k, and verifier best-of-N are reported separately. No benchmark-specific tuning.
- Clean data. We use commercially licensed data only, with no distillation from closed models.
Roadmap
- DAMS-2: a mixture-of-experts upgrade aimed at being the best open coding model.
- DAMS-3: much larger-scale verified agentic RL, verifier-guided search, recurrent depth, and persistent repo memory.
- DAMS-4: a fully custom Oriven-native architecture trained from scratch.
Contact
alexander@orivenlabs.com