read · chart · name
READ each open model's nature, CHART its descent from base to finetune to merge, and NAME the drift along every branch — an honest, cited tree of who-came-from-whom.
82 specimens · 6 family groves · 56 lines of descent · 2 merges · 8 inferred · 7 read in the atlas
charted 2026-07-07 · a curated spanning set, not a census · transcribed from the research atlas
the tree of life
A handful of base pretrains — Llama, Mistral, Qwen, Gemma, Phi, DeepSeek — carry almost the whole canopy. The base stock sits at the headwaters at the top; descent flows downward through finetunes and merges, one family per grove column. Where two branches rejoin — a frankenmerge, or a distillation grafting a teacher onto a different body — the river braids. Pick a grove to frame its family, drag to pan, zoom with ctrl + scroll (or a pinch, or a double-click), and use Fit or fullscreen to frame the whole canopy — a plain scroll stays with the page.
the lines of descent
the groves
the training-stage moves
generational succession — the maker lines
A new generation (Llama 2 → 3, Qwen 2 → 2.5, Phi-2 → Phi-3) is a fresh pretrain, not a weight-descendant of the last — so these are kept OUT of the lines of descent above. A generation hop is never mistaken for weight inheritance.
Meta Llama · Llama
Each Llama generation (1 -> 2 -> 3 -> 3.1) is a FRESH pretrain, not a weight-descendant of the last; the exception is Llama 3.2's small models, which ARE weight-derived (pruned+distilled) from 3.1 — so that one hop is a real edge in the tree, the rest are generational succession only.
Mistral AI · Mistral / Mixtral
Mistral-7B, the Mixtral MoEs, and Mistral-Nemo are each independent fresh pretrains — Mixtral is NOT an upscale of the 7B. (SOLAR-10.7B is the one model that DOES weight-inherit from Mistral-7B, via depth up-scaling — a real edge.)
Alibaba Qwen · Qwen
Qwen 1.5 -> 2 -> 2.5 -> 3 are successive fresh pretrains; the WITHIN-2.5 specialisations (Instruct, Coder, Math) ARE weight-derived edges.
Google Gemma · Gemma
Gemma 1 -> 2 -> 3 are fresh pretrains built on shared Gemini research; WITHIN Gemma 2 the small models are distillation-derived from the 27B teacher (real edges).
Microsoft Phi · Phi
The whole Phi line is generational: each is a fresh synthetic-data ('textbooks') pretrain on an evolving recipe, NOT a weight-descendant of the previous Phi. There are no intra-Phi weight edges in this tree — the shared inheritance is the DATA PHILOSOPHY, not the weights.
DeepSeek · DeepSeek
DeepSeek-LLM, Coder, and V3 are independent fresh pretrains; R1 IS weight-derived from V3-Base (a real edge), and the R1-Distill models graft R1's reasoning onto Qwen/Llama bodies (dual-parent edges).
Every node that Ardora has read carries a full field-note in the atlas — its Silhouette, its abstains, and the change a finetune made off its champion, read as a contrast.
