Deem 0.6B (v1)
Decisions, anywhere — smaller. The Deem decision stack on CPU at 0.6B parameters: typed, calibrated choices, scores, and yes/no decisions with abstention, served by our own Rust runtime. No GPU required.
Why 0.6B?
Qwen3-0.6B-Base has dense attention where its Qwen3.5 successor hybridizes with linear attention. Through the Deem letter-slot fine-tune that matters: 0.6B beats 0.8B on the standard suite (0.7533 vs 0.7455 macro) and on JevBench hard (40.5 vs 33.3) — with 25% fewer parameters.
Benchmarks
- 94.5% long-policy hold-out accuracy
- Standard-suite macro 0.7533 (0.7622 calibrated)
- JevBench public 97.9 / 73.6 / 40.5 (easy / original / hard)
- Near-ceiling exact-law reasoning: counting 0.988, grid 0.992, zero-count 1.000 — best exact-laws Brier on the Tare board (0.0289)
- Temporal probes match the 9B: sharp 30/31-day window cutoff, consistency sums 1.000
- 65ms short-form decisions, 1.9s ~3k-token states (int8 path, busy desktop CPU)
- ~0.7GB resident (int8 path)
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 - Parity-gated against torch: max letter-logit diff 0.080 bf16 / 0.125 int8 (gate 0.35)
- 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-0.6-v1 ./target/release/deem-server
Runs where GPUs don't: CI runners, edge boxes, laptops, serverless micro-VMs.
How it's built
Qwen3-0.6B-Base (Apache-2.0, license tag + LICENSE file verified on HF 2026-09-27), full fine-tune on the Deem ground-truth-verified mixture (124,765 rows: anchors + exact laws
- policy + working-time + verification traces), then the temporal-window delta (6k windowgen + 15k replay, lr 1e-5). Apache-2.0 recipe.
License
Apache-2.0. All benchmarks reproducible from the release artifacts.
- Downloads last month
- 17