Fine-tunes & community model names
HuggingFace is full of community models with long, keyword-stacked names. The pattern is always the same: base model + size + what was done to it + how it’s packaged. The vocabulary:
Base model — The original model everything else is derived from. Community models almost always name their base first, because that determines architecture, tokenizer, and rough capability.
Open-weight vs. fully open — “Open-weight” means only the trained weights are published (Llama, Qwen, Mistral) — you can run and fine-tune the model but not reproduce it. “Fully open” (Olmo, Apertus, Soofi) additionally releases the training data recipe, training code, and intermediate checkpoints, so a third party can audit or rebuild the run. Model cards and papers use the distinction precisely; marketing often blurs both into “open source”.
Sovereign model / sovereign AI — A model trained end-to-end on domestic infrastructure under local (usually European) data-protection and operational rules, typically with publicly documented data — e.g. Apertus (Switzerland), Teuken/Soofi (Germany), EuroLLM. The label is about who controls the training pipeline and where it ran, not about architecture or quality.
Fine-tune / finetune — Taking a base model and training it further on extra data to change its behavior (coding, roleplay, a specific style). Cheap compared to training from scratch; the vast majority of HuggingFace uploads are fine-tunes. Flavors you’ll see in cards: SFT (supervised fine-tuning — train on example conversations), DPO/RLHF/GRPO (preference tuning — train on “this answer is better than that one”).
LoRA / QLoRA — The cheap way to fine-tune: instead of updating all weights, train a small “adapter” bolted onto them (QLoRA = doing that on top of a quantized model, so it fits on consumer GPUs). Adapters can be shipped separately (a few hundred MB) or merged into the weights before upload.
Instruct / -it / chat — A model tuned to follow instructions in a conversation (vs. a raw “base”/“text” model that just continues text). Nearly everything you’d self-host for chat is an instruct model.
Distillation (distill) — Training a smaller/cheaper model on the outputs of a bigger model so it imitates it. A bigger model’s name inside a fine-tune’s name (e.g. “R1-Distill”, or a Claude/GPT model name) means “trained on data generated by that model”, not that it contains that model’s weights.
Merge / model merging — Combining the weights of several fine-tunes of the same base into one model (no training involved, just arithmetic). Popular in the hobbyist scene; quality is hit-or-miss. The standard tool is mergekit; method names that show up in cards: SLERP, TIES, DARE / DARE-TIES, Model Stock, task arithmetic — with per-model weight and density knobs controlling how much each ingredient contributes.
Codenames & purpose words in names — Fantasy words in a name are pure branding — the uploader’s name for their recipe, with no standard meaning. Functional words hint at the target use: Novelist / Writer / Storyteller = creative prose, Coder = programming, RP = roleplay, Agentic = multi-step tool use (read → reason → act loops, as in coding agents). Tier words borrowed from big-vendor naming (mini / nano / flash / Turbo) suggest a smaller or faster variant but carry no standard meaning on community models. A Max suffix on vendor models often means the same model run at maximum reasoning-effort settings, not a different architecture. Version tags (v0.2, -preview) and context tags (-128K) mean what they say.
Benchmark names & scores in names — A benchmark name in a model name (tau2 = τ²-bench, MMLU, arena) means the fine-tune targets or advertises that benchmark. Self-reported multipliers (“3.5x”) compare against a baseline the uploader chose, on the benchmark the model was tuned for — treat them as marketing until independently reproduced.
Uncensored / abliterated / abliteration — The model’s refusal behavior (“I can’t help with that”) has been removed. Abliteration is the common technique: find the internal direction that represents refusal and delete it, no retraining needed. Heretic is a popular open-source tool that automates this. Side effect: can make models dumber or less stable.
Frankenmodel / upscaling / parameter expansion — Making a model bigger than its base by duplicating/stacking layers, then (sometimes) healing it with more training (e.g. a 27B base sold as “40B”). The extra parameters are not new knowledge — treat capability claims skeptically.
Thinking / Reasoning (in a name) — The fine-tune emits chain-of-thought before answering (see CoT). Often paired with a base that supports it natively.
RP / roleplay models — Fine-tunes optimized for creative writing/character roleplay (a large share of the merge/uncensored scene).
imatrix (importance matrix) — Calibration data used when making GGUF quants (the basis of IQ-quants): sample text is run through the model to measure which weights matter most, so those keep more precision. “IMatrix” in a repo name = the quants were made with one; the choice of calibration text is up to whoever made the quants. Other quant-related keywords in such names (custom dataset names, “MAX”, etc.) are usually the uploader’s own branding for their recipe, not standard terms.
-GGUF suffix / quant repos — A repo name ending in -GGUF is not a new model: it’s someone’s repackaging of an existing model into GGUF quant files for llama.cpp/Ollama (one file per quant level, Q4_K_M etc. — see Quantization). The same pattern exists for other runtimes and formats: -AWQ / -GPTQ / -exl2 (GPU quant formats), -MLX (Apple Silicon), -ONNX, or “for
Rule of thumb: the longer the name, the further the model is from anything validated. Base-vendor releases and QAT checkpoints have benchmarks; ten-keyword merge/uncensored/frankenmodel stacks usually have only the uploader’s word.