LibertAIDAI/deem-0.8-v1

Deem 0.8B (v1.1)

Decisions, anywhere. The full Deem decision stack on CPU — typed, calibrated choices, scores, and yes/no decisions with abstention, served by our own Rust runtime. No GPU required.

Update (v1.1): this release improves accuracy on temporal comparisons. v1.0 answered compact return-window questions ("order 60 days old" — "within the 30-day window?") from a yes-prior, and could rate P(within) + P(outside) up to 1.9. v1.1 trains the comparison itself across phrasings, with faithful verification traces: return-policy accuracy rises to 98.2% (from 96.6%), the window boundary is sharp (30 in / 31 out), and polarity consistency is exact — P(within) + P(outside) = 1.0 at every age. Standard-suite and hold-out scores hold steady. The v1.0 weights remain available under v1.0/.

Benchmarks

• 96.2% long-policy hold-out accuracy (trap/adversarial items: 91.5%)
• 98.2% returns-domain accuracy (v1.0: 96.6%)
• 362ms short-form decisions on a busy desktop CPU
• 0.9GB resident (int8 path; 1.6GB bf16-exact)
• Near-ceiling exact-law reasoning: counting 0.976, grid 0.984, zero-count 1.000

The runtime

A single static Rust binary. No Python, no C++ dependencies at runtime.

• Hand-written AVX-512 kernels — bf16 (vdpbf16ps) and int8 (vpdpbusd) GEMM lanes, ~3.4 TFLOPs/s isolated int8
• Parity-gated against torch: max letter-logit diff 0.085 (bf16 noise floor)
• Wire-compatible /v1/systemone — drop-in for the TypeSafe SDK

git clone https://github.com/Libertai/deem && cd deem/rust
cargo build --release
DEEM_CHECKPOINT=LibertAIDAI/deem-0.8-v1 ./target/release/deem-server

Runs where GPUs don't: CI runners, edge boxes, laptops, serverless micro-VMs.

License

Apache-2.0 (weights, code, and benchmark).