this-that-model

Last updated:

Open20–80ms$0/M input

Quadrant scores

See the full quadrant

Scored with the public quadrant rubric (maturity × capability; bubble = adoption). Revise as evidence lands.

  • Maturity4.2/10

    Day-old MIT weights + inference repo and a GGUF mirror, adapted from Mapika/decider-2b; no hosted SLA and almost no HF traction yet keep maturity early-mid.

  • Capability4/10

    Decision Index 0.3.2 Full 24.56 (#130; board row 'this-that 1.2'). Typed choice over declared options in one pass; probabilities described as calibrated but no ECE published; ~30.9 ms/q is the author's. The 68q table and spatial bench are UnverifiedClaim.

  • Adoption12/100

    FLock X announce (~53♥ / 10 RT on JUST IN) plus HF card (~1♥ / ~86 downloads) and ~3 GH stars — visible in the open decision wave, still niche.

Vendor claims

News covering this model

Board audit of 17 Oct 2026: see the “Board audit” section below; where it conflicts with older board statements on this page, it wins.

this-that-model is an open typed decision model from FLock.io — ~1.88B hybrid Qwen3.5 / DeltaNet weights (MIT), adapted from Mapika/decider-2b under Apache-2.0 — not a TypeSafe product and not a logit-wrapper method-only trick.

One forward pass returns an index over options you declare plus calibrated probabilities; zero output tokens, no decode loop, no parser. Inference: github.com/FLock-io/this-that-model. Weights: flock-io/this-that-model-1.0. Quantized mirror: mradermacher/this-that-model-1.0-GGUF. Spatial eval set: limberc/this-that-spatial-bench.

FLock announced the release on X on 21 Sep 2026 (@flock_io).

Limits (read before you ship)

  • Author 68-question head-to-heads (incl. vs Jev / GPT-5.6) and ~30.9 ms/q latency are self-reported → UnverifiedClaim; ModelSystem.One has not reproduced them
  • Card itself flags search / multi-step arithmetic ceilings and English-only short states — plan around those
  • Catalog decisionTypes list choice only (declared options); no first-class TypeSafe score / noul wire API here
  • Different contract than hosted Jev or sibling Decider System One JSON — do not drop-in swap without an adapter

Why it is in the catalog

Public trained typed-decision weights (adaptation of Decider-2B under a proper scoring rule), open MIT release + inference code, and a clear System One peer next to Decider / NanoJev — not a frozen-logit method-only wrapper.

Board audit (17 Oct 2026)

Checked against the official boards on 17 Oct 2026: Decision Index 0.3.2 (data of 16 Oct, 19:00 UTC), JevBench and ImageJevBench live. The boards are separate sources and are never averaged. Older board statements and rankings further down this page are superseded where they conflict with this section.

Board Row Result
Decision Index 0.3.2 (16 Oct 19:00 UTC) FLock this-that 1.2 Full 24.56 (#130, tie group 36); public 28.65 (#133); same-skills 27.51; new-domains 19.7

Score proposal (from scratch, audit follow-up 17 Oct 2026). Capability 6.0 → 4.0 = types × 0.30 + calibration × 0.25 + latency × 0.25 + accuracy × 0.20 = 3 × 0.30 + 4 × 0.25 + 6 × 0.25 + 3 × 0.20 = 4.00 → 4.0. Types 3: decisionTypes lists choice only (the earlier ~9 counted the schema, not the catalog types). Calibration 4: probabilities claimed with a proper scoring rule but no ECE or independent study. Latency 6 [vendor]: ~30.9 ms/question on a consumer GPU, single setup, no p99. Accuracy 3: Decision Index Full 24.56 ÷ 60.11 × 8 = 3.3 → 3. Caution: the board row is named ‘this-that 1.2’ while this page is the 1.0 release; the version match is not confirmed. Maturity and adoption unchanged.

Hugging Face on 17 Oct (flock-io/this-that-model-1.0): 1,681 downloads, 18 likes, licence tag mit, last modified 2026-09-21.