FLock ships this-that-model — open typed decisions in ~31 ms

Source · Model card: this-that-model

On 21 September 2026, FLock.io announced this-that-model-1.0 on X (@flock_io): an open typed decision model that returns one of n declared options plus calibrated probabilities in a single forward pass — no decode loop, zero output tokens.

The card lists ~1.88B parameters on a Qwen3.5-style hybrid stack (18/24 layers DeltaNet linear attention, 6 full attention), MIT licence, and inference at github.com/FLock-io/this-that-model. Training note: adapted from Mapika decider-2b (Apache-2.0) under a strictly proper scoring rule. A community GGUF mirror is up at mradermacher/this-that-model-1.0-GGUF; spatial eval data at limberc/this-that-spatial-bench.

UnverifiedClaim / hygiene: the author 68-question table (this-that 0.941 vs Jev 0.765 vs GPT-5.6 0.926) and the ~30.9 ms/question median on a consumer GPU are self-reported from the HF card — not ModelSystem.One benches. The card itself flags search / multi-step arithmetic limits and English-only short states.

Catalog entry: /models/this-that. Peer context: Decider, NanoJev, hosted Jev.