# Decision models are starting to have family trees

> Fine-tunes now stack on fine-tunes. REAX 1.1 descends from Perplexity's pplx-decider, Torchcast from StartLux, and a recommender reranker from Kev-9B. Licences and quirks travel with them.

Published 2026-10-23

Canonical: https://shipped.anth.us/stories/decision-models-family-trees/

Decision models read a state and a list of questions. They return a probability for each answer and write no prose. They are new, and a family tree is already visible.

The longest branch runs through Perplexity. REAX 1.1 is a fine-tune of REAX 1, which is a fine-tune of pplx-decider-v1.2, derived from Qwen3.8-27B [1]. On the card's own scoring the three step up from 65.34 to 67.65 to 71.27. REAX says it is not affiliated with Perplexity. Perplexity's own page covers v1 and v1.1, both Apache-2.0 [2]. The v1.2 link rests on REAX's card alone.

The second branch runs through StartLux. Torchcast Decision 27B is StartLux-Decision-27B with a LoRA fine-tune merged in [3]. The card reports 65.00 against 63.88 for the parent, a gain of 1.12 points on its own index. The weights inherit the parent's CC BY-NC 4.0 licence, and StartLux's later v1.1 family keeps it [4]. A child is no freer than its parent.

The third branch is the smallest. Kev-9B is a rank-16 LoRA adapter and a pointer head on Qwen3.5-9B-Base [5]. Decision4Rec starts from that adapter and trains it to rerank recommender candidates [6]. Its results are development numbers only.

Some look-alikes are not relatives. Matilda Decision v1 is post-trained from Qwen3.8-27B directly [7]. So is the vLLM team's Vega-27B [8]. Neither page names pplx-decider as a parent.

Two roots carry much of the tree. Qwen3.5 sits under Amazon's 2B decider [9] and H2O's 4B [10]. Qwen3.8-27B sits under Perplexity [2], Matilda and Vega.

The kinship is also in the interface. Kev implements TypeSafe's System One API [5]. Torchcast, H2O and Amazon serve the /v1/systemone route, and Torchcast's card says clients written for Jev should work unchanged [3][10][9].

Traits travel too. StartLux warns that engines reading the prompt causally serve a different model [4]. Perplexity and Matilda say the same about attention [2][7]. Families pass on licences, limits and habits. Read the parent's card before the child's.