LibertAI Labs
← All models

Model

Shipped

Deem 9B

The strongest open decision model we know of. Typed, calibrated decisions on long state, from a letter-slot readout, in one prefill pass.

Base model:
Qwen/Qwen3.5-9B-Base
License:
Apache-2.0
JevBench hard (composite)
65.8
single-pass 54.1, extended 68.9
JevBench easy / original
100.0 / 91.7
Judge-tier verification
16/17
Long-state P50
788ms
3,222-token state, edge GPU
Short-form P50
103ms

Deem 9B reads your state (a policy, contract, ticket, or question) and returns a typed, calibrated decision: Choice over 2 to 255 options, Score on 2 to 10 ordered levels, or Noul (yes/no) with abstention. One forward pass, prefill-only, no decode phase.

The shipped configuration is confidence-gated: a direct pass always runs, and a chain-of-thought pass triggers when the direct readout’s entropy exceeds a threshold. 58% of items resolve in a single pass; the rest pay for extended reasoning. The composite JevBench numbers on the model card use tier routing (direct on easy, extended elsewhere); every mode is stated with its number, and the single-pass figure is 54.1 hard, published next to the composite.

Every training domain is ground-truth verified: generator-built states with labels by construction, including GenRM-style verification traces that took a judge yes-bias from 9/17 to 16/17 in one training cycle. Measured on JevBench public (231 items), never trained on. Apache-2.0 recipe, start to finish.