LibertAI Labs
← All models

Model

Shipped

Deem 0.8B

The full Deem decision stack on CPU. Typed, calibrated choices, scores, and yes/no decisions served by our own Rust runtime. No GPU required.

Base model:
Qwen/Qwen3.5-0.8B
License:
Apache-2.0
Long-policy hold-out
96.3%
trap/adversarial items 91.8%
Short-form latency
362ms
contended desktop CPU
Resident memory
0.9GB
int8 path, 1.6GB bf16-exact
Exact-law zero-count
1.000
counting 0.976, grid 0.984

Decisions, anywhere. Deem 0.8B is the CPU-native sibling of Deem 9B: the same typed, calibrated decision stack on a Qwen3.5-0.8B full fine-tune over the ground-truth-verified mixture, served by a single static Rust binary. No Python, no C++ dependencies at runtime.

The runtime carries hand-written AVX-512 kernels, bf16 (vdpbf16ps) and int8 (vpdpbusd) GEMM lanes, and is parity-gated against PyTorch to the bf16 noise floor (max letter-logit diff 0.085). The wire format is /v1/systemone, drop-in compatible with the TypeSafe SDK. It runs where GPUs do not: CI runners, edge boxes, laptops, serverless micro-VMs.

Its external hard-split number is 33.3, stated plainly; its strength is the long-policy domain, where 96.3% hold-out accuracy and 91.8% on trap items make it the deployable one. All benchmarks are reproducible from the release artifacts.