Deem 9B (v1)
LibertAI's open-weights decision model on Qwen3.5-9B-Base. It reads a state and returns a typed choice, ordinal score or yes/no with abstention, read from a letter slot in one prefill pass with a probability per option. CPU-native 0.8B and 0.6B siblings.
A LoRA-merged fine-tune with a letter-slot readout; the served response shows zero output tokens. Siblings LibertAIDAI/deem-0.8-v1 (Qwen3.5-0.8B; root weights are v1.1, with v1.0 kept in a subfolder) and LibertAIDAI/deem-0.6-v1 (Qwen3-0.6B-Base) run on LibertAI's own Rust CPU runtime with AVX-512 bf16 and int8 kernels; the maker reports 96.2% on its long-policy hold-out for the 0.8B, at 362 ms per short decision on a busy desktop CPU. On the third-party JevBench public set the maker reports, for the 9B, 100.0 easy, 91.7 original and 65.8 hard (68.9 with its optional extended-reasoning mode, which generates text before answering and is not one pass). LibertAI's servers cap choice questions at 26 options although the card says 2 to 255. Temperatures are shipped but no calibration figure is published. The maker says it never trained on JevBench. No hosted API. Not related to mertkayacs/Deem-4B.
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 probability
At a glance
| Parameters | 9B |
| Base model | Qwen/Qwen3.5-9B-Base |
| Maker | LibertAI |
| Released | 2026-09-24 |
| License | apache-2.0 |
| Reported accuracy | 65.8% |
| Reported latency | about 100 ms p50 per short decision on a low-power edge GPU (model not named); 788 ms p50 for policies of 3,200+ tokens |
Get the weights
pip install systemonemodels
systemone pull libertai/deem
The files are served from the maker's Hugging Face repository, LibertAIDAI/deem-9b-v1, and verified against the checksums recorded here.