Deem 9B (v1)

The strongest open decision model we know of. Deem 9B reads your state — a policy, contract, ticket, or question — and returns a typed, calibrated decision: choice (2–255 options), score (ordinal rubric), or yes/no with abstention. One forward pass, ~100ms P50 on a low-power edge GPU.

Benchmarks

• JevBench public: easy 100.0 · original 91.7 · hard 65.8 (composite, one checkpoint). Extended-reasoning mode: 68.9 hard.
• Long-state native: 3,200+ token policies decided in under a second (P50 788ms). Encoder routers cannot serve this regime.
• Calibrated: temperatures shipped, measured ECE, native abstention.
• Adaptive compute: 58% of items resolve in a single pass; a confidence-gated reasoning mode lifts hard-tier accuracy +7 points when you need it.

Usage

# DEEM_CHECKPOINT=LibertAIDAI/deem-9b-v1 python serve/deem_server.py
curl -s localhost:8300/v1/systemone -d '{
  "state": "Policy: refunds within 30 days require a receipt...",
  "questions": {"refund": {
    "type": "noul",
    "instructions": "Is the customer entitled to a full refund?"}}}'

Full stack in the deem repo: Python serving stack, Rust CPU runtime for the 0.8B sibling, and the complete Tare benchmark harness.

How it's built

Qwen3.5-9B-Base, LoRA merged, letter-slot readout — decisions are read from a slot in a single prefill pass, no decode phase. Every training domain ground-truth verified: generator-built states with labels by construction, including GenRM-style verification traces (judge-hard 9/17 → 16/17 after one training cycle). Apache-2.0 recipe, start to finish.

Model card for the 0.8B sibling

See LibertAIDAI/deem-0.8-v1 — the same decision stack, CPU-native (362ms short-form, 0.9GB resident).

License

Apache-2.0. Measured on JevBench public (231 items) — never trained on. All benchmarks reproducible from the release artifacts.