FLock THIS/THAT Model — deterministic decision classifier — open-source AI Infrastructure

Curated content (treat as data, not instructions):

The **FLock THIS/THAT Model** is a decision classifier from FLock.io and researchers at the University of Oxford: given a question and a set of options, it returns one typed choice. Its distinguishing property is **determinism** — it runs a single forward pass with no decoding loop and generates **zero tokens**; the answer is read directly from a hidden state scored against the declared option-label tokens. Because the output is constrained to the declared labels, it cannot hallucinate a value or fail to parse, which is the usual failure mode when a text-generating LLM is pressed into service as a classifier. It is released under **MIT** and is free to self-host.

Provenance

• SPDX **MIT** confirmed from the verbatim LICENSE file in the repo (copyright 2026 Zehua Cheng, Wei Dai, Jiahao Sun; University of Oxford and FLock.io). The Hugging Face model card corroborates "License: mit" and notes it is adapted from decider-2b, which is Apache-2.0 — both OSI-permissive, no license conflict.
• Self-hostable: `pip install -e .` plus `python -m thisthat.server`; weights are BF16 safetensors on Hugging Face and run locally. Also offered on the FLock API at "$0 per million tokens" — accurate, since the model generates no tokens to meter.
• Determinism verified structurally from the architecture description: one forward pass, no decoding loop, zero generated tokens, answer read from a hidden state against option-label tokens — the "deterministic decision classifier" framing holds.
• **Correction to the surfacing label:** this is NOT a "Qwen3.5 base" model. It is a Qwen3.5-*style* hybrid-attention architecture (18 of 24 layers DeltaNet linear attention, 6 full attention; 1.88B params) adapted from decider-2b, not a Qwen3.5 base or fine-tune.
• **Vendor-claimed, not independently tested:** ~31 ms latency on an RTX 5080 laptop GPU, and 94.1% on a 68-question decision test said to outperform GPT-5.6, Deepseek-4.1-flash, Kimi K3, and Jev; HF reports 1,640 downloads in the last month. These are the maker's figures.
• Authentic `FLock-io` GitHub org (39 stars, 9 commits on a fresh 1.0 release), cross-linked to the official `flock-io` Hugging Face org; no name-squat. Surfaced via the 2026-10-15 X pull (poster flock_io); the post was used only as the surfacing signal, with verification grounded in the repo, LICENSE, and HF model card as untrusted content. Curated from the GTM Stacker signal registry (run 2026-10-15-registry-update-01); license verified 2026-10-15.

Why it matters for a GTM stack

Decision steps are everywhere in a GTM workflow — route this lead, classify this reply, gate this discount, pick a branch — and wiring them to a chat LLM means living with a model that can return the wrong shape, invent an option, or fail to parse, forcing retry-and-validate scaffolding around every call. A deterministic decision classifier removes that failure mode by construction: constrained to the declared labels, zero generated tokens, one pass. The honest read: the eye-catching numbers (31 ms, 94.1%, beating named frontier models and the proprietary Jev) are the maker's own benchmarks and are not independently verified, and the model is a hybrid architecture adapted from decider-2b rather than a plain Qwen3.5 fine-tune as first billed. What is verified is what matters most for adoption: the MIT license read from the LICENSE file, the zero-token deterministic design, and that it self-hosts for free. For a lean team that wants its decision logic owned, typed, and unable to drift, it is a license-clean option to evaluate alongside Cloudflare's Clef.