Supported models

Every model below can be passed as the model field when creating an AutoScientist run. The runtime model list is authoritative; use this page as a planning reference.

response = client.autoscientist.list_models()
for model in response.models:
    print(model.id, model.api_model_size, model.methods, model.training_types)

See the autoscientist.list_models API reference for the current response schema.

You do not need to select a model. Omit model and AutoScientist selects one that supports the requested training_method. The resolved model ID is returned on the run.

Available models

Model	model ID	Size	Methods	Supported training_type	Supports reasoning training	Min. rows
SFT / Alignment
Gemma 3 (4B Instruct)	google/gemma-3-4b-it	4B	SFT, DPO	lora, full	false	1,000
Gemma 4 (31B Instruct)	google/gemma-4-31B-it	31B	SFT, DPO	lora, full	true	10,000
gpt-oss (20B)	openai/gpt-oss-20b	20B	SFT, DPO	lora	true	1,000
gpt-oss (120B)	openai/gpt-oss-120b	120B	SFT, DPO	lora	true	1,000
Llama 3.3 (70B Instruct)	meta-llama/Llama-3.3-70B-Instruct-Reference	70B	SFT, DPO	lora, full	false	1,000
Llama 4 Scout (17B-16E Instruct)	meta-llama/Llama-4-Scout-17B-16E-Instruct	109B	SFT, DPO	lora	false	1,000
Meta Llama 3.2 (3B Instruct)	meta-llama/Llama-3.2-3B-Instruct	3B	SFT, DPO	lora, full	false	1,000
Mistral (7B) Instruct v0.2	mistralai/Mistral-7B-Instruct-v0.2	7B	SFT, DPO	lora	false	1,000
Mixtral 8x7B Instruct (v0.1)	mistralai/Mixtral-8x7B-Instruct-v0.1	46.7B	SFT, DPO	lora, full	false	1,000
NVIDIA Nemotron 3 Nano Omni 30B A3B Reasoning BF16	nvidia/Nemotron-3-Nano-Omni-30B-A3B-Reasoning-BF16	30B	SFT, DPO	lora	true	10,000
NVIDIA Nemotron 3 Super 120B A12B BF16	nvidia/NVIDIA-Nemotron-3-Super-120B-A12B-BF16	120B	SFT, DPO	lora	false	10,000
Qwen 3.5 (0.8B)	Qwen/Qwen3.5-0.8B	0.8B	SFT, DPO	lora	false	1,000
Qwen 3.5 122B A10B	Qwen/Qwen3.5-122B-A10B	122B	SFT, DPO	lora	true	10,000
Qwen 3.6 35B A3B	Qwen/Qwen3.6-35B-A3B	35B	SFT, DPO	lora	true	10,000
SFT / Alignment (multimodal)
Gemma 3 (4B) VLM	google/gemma-3-4b-it-VLM	4B	SFT, DPO	lora	false	1,000
Gemma 3 (27B) VLM	google/gemma-3-27b-it-VLM	27B	SFT, DPO	lora	false	1,000
Gemma 4 (31B) VLM	google/gemma-4-31B-it-VLM	31B	SFT, DPO	lora	false	1,000
Qwen 3.5 (4B)	Qwen/Qwen3.5-4B	4B	SFT, DPO	lora	true	1,000
Qwen 3.5 (9B)	Qwen/Qwen3.5-9B	9B	SFT, DPO	lora	false	1,000

Min. rows is the minimum effective dataset size (base rows plus augmentation) required to select that model. DPO training requires at least 12,000 rows regardless of model; see Row count.

A full value for training_type means full fine-tuning is available in addition to lora.

Per-model training limits

These limits are properties of the model and training hardware. AutoScientist accounts for them automatically; they matter primarily when you override its recommended hyperparameters.

model ID	Context (SFT)	Context (DPO)	Max batch (SFT)	Max batch (DPO)	Gradient accumulation	Maximum lora_r
SFT / Alignment
google/gemma-3-4b-it	131072	65536	8	8	1	64
google/gemma-4-31B-it	49152	24576	4	4	2	64
openai/gpt-oss-20b	131072	65536	1	1	8	64
openai/gpt-oss-120b	65536	32768	2	2	8	64
meta-llama/Llama-3.3-70B-Instruct-Reference	24576	12288	8	8	1	64
meta-llama/Llama-4-Scout-17B-16E-Instruct	65536	12288	8	8	1	64
meta-llama/Llama-3.2-3B-Instruct	131072	65536	8	8	1	64
mistralai/Mistral-7B-Instruct-v0.2	32768	32768	16	8	1	64
mistralai/Mixtral-8x7B-Instruct-v0.1	32768	16384	8	8	1	64
nvidia/Nemotron-3-Nano-Omni-30B-A3B-Reasoning-BF16	65536	32768	8	8	1	64
nvidia/NVIDIA-Nemotron-3-Super-120B-A12B-BF16	49152	24576	2	2	4	64
Qwen/Qwen3.5-0.8B	131072	131072	8	8	1	64
Qwen/Qwen3.5-122B-A10B	65536	32768	16	16	1	64
Qwen/Qwen3.6-35B-A3B	65536	32768	8	8	1	64
SFT / Alignment (multimodal)
google/gemma-3-4b-it-VLM	32768	32768	8	8	1	64
google/gemma-3-27b-it-VLM	32768	24576	8	8	1	64
google/gemma-4-31B-it-VLM	24576	12288	8	8	1	64
Qwen/Qwen3.5-4B	131072	65536	8	8	1	64
Qwen/Qwen3.5-9B	65536	49152	8	8	1	64

Batch size

Each model has one valid batch size. batch_size defaults to "max", which resolves to that model-specific value. Leaving it unchanged is the right choice for almost every run. Supplying another integer causes the launch to fail after the job is accepted.

Models use gradient accumulation when a larger effective batch is useful.

# Recommended: let AutoScientist resolve the model and batch size.
run = client.autoscientist.create(dataset_id=dataset_id)

# Advanced: pin a model and its required batch size.
run = client.autoscientist.create(
    dataset_id=dataset_id,
    model="openai/gpt-oss-20b",
    hyperparams={"batch_size": 1},
)

LoRA rank

lora_r accepts values from 1 through 64 for every model on this page. lora_alpha must be exactly one or two times lora_r; the API cross-checks the values when both are supplied in the same request.

Row count

AutoScientist enforces the following minimum effective dataset sizes (base rows + augmentation):

Training method	Minimum rows	Applies to
SFT (instruction)	1,000	All models
DPO (preference)	12,000	All models
Any method	10,000	Gemma (≥ 26B), Qwen (≥ 35B), NVIDIA

Text datasets can use augmentation to reach a minimum. Multimodal datasets cannot currently be augmented, so they must already contain enough processed rows. AutoScientist treats 20,000 rows as a recommended quality target, not a hard launch requirement.

Context length and your data

Training context is the maximum sequence length used during fine-tuning. It is shorter than the serving context. Rows longer than the training limit are truncated rather than rejected.

DPO context can be shorter than SFT context. Before pinning a model for DPO, confirm that DPO appears in its runtime methods list and use the DPO context column above.