Supported models
Load one model per session. Catalog slugs (gpt2-small, gte-small) include curated public SAEs where available. Family support lets you pass a HuggingFace repo id for other sizes in the same architecture — GPT-2, Pythia, Llama, and common embedding encoders. Run aquin list model for the catalog and family examples.
Catalog models
Curated slugs with public SAE checkpoints where available.
LLMs
Inspection, simulation, evals, and checkpoint analysis.
- GPT-2 SmallLLM
gpt2-smallPublic Neuronpedia SAE - Pythia 70MLLM
pythia-70mPublic HF SAE - Pythia 2.8BLLM
pythia-2.8b - Llama 3.2 1BLLM
llama-3.2-1b - Sarvam 30BLLM
sarvam-30b
Embedding models
Layer drift, retrieval, embed simulation, and embedding evals. See Inspection (embedding).
- all-MiniLM-L6Embedding
all-minilm-l6 - GTE SmallEmbedding
gte-small - BGE Small ENEmbedding
bge-small-en
Model families
Pass any example HuggingFace id to aquin session start --model. Aquin introspects layer count and width, then runs the same tool suite as the catalog entry. Public SAE pull (aquin load sae) applies to catalog models; on family models use aquin sae train to build a temporary SAE first.
LLM families
GPT-2
OpenAI GPT-2 sizes. Inspection and checkpoint tools; train your own SAE with aquin sae train.
gpt2 · gpt2-medium · gpt2-large · gpt2-xl
Pythia
EleutherAI Pythia checkpoints. Pass the full HuggingFace repo id to --model.
EleutherAI/pythia-160m-deduped · EleutherAI/pythia-410m-deduped · EleutherAI/pythia-1.4b-deduped · EleutherAI/pythia-2.8b
Llama
Meta Llama instruct checkpoints (gated — aquin login HF proxy). Catalog llama-3.2-1b includes a native public SAE.
meta-llama/Llama-3.2-1B-Instruct · meta-llama/Llama-3.2-3B-Instruct · meta-llama/Llama-3.1-8B-Instruct
Embedding families
BERT-style encoders
Sentence-transformers, GTE, BGE, and E5 families. Full embed inspection suite after session start.
thenlper/gte-base · BAAI/bge-base-en-v1.5 · sentence-transformers/all-mpnet-base-v2 · intfloat/e5-small-v2
SAEs & checkpoints
- One model per session. Start a new session to switch.
- Catalog: feature tools need a public SAE — aquin list sae, then aquin load sae <model-l{n}>.
- Family models: train with aquin sae train, then aquin load sae --user <name> before inspect.
- Fine-tuned weights stay on disk — use --checkpoint while the base model id stays loaded.
- SAE catalog:
aquin list sae· SAE training · Checkpoint tools
