Supported models
Base models available to AutoScientist and the per-model limits for context length, batch size, and LoRA rank.
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.
Available models
Section titled “Available models”| Model | model ID | Size | Methods | Training types |
|---|---|---|---|---|
| Qwen 3.5 (0.8B) | Qwen/Qwen3.5-0.8B | 0.8B | SFT, DPO | LoRA |
| Meta Llama 3.2 (3B Instruct) | meta-llama/Llama-3.2-3B-Instruct | 3B | SFT, DPO | LoRA, full |
| Gemma 3 (4B Instruct) | google/gemma-3-4b-it | 4B | SFT, DPO | LoRA, full |
| Gemma 3 (4B) VLM | google/gemma-3-4b-it-VLM | 4B | SFT, DPO | LoRA |
| Mistral (7B) Instruct v0.2 | mistralai/Mistral-7B-Instruct-v0.2 | 7B | SFT, DPO | LoRA |
| Qwen 3.5 (9B) † | Qwen/Qwen3.5-9B | 9B | SFT, DPO | LoRA |
| gpt-oss (20B) | openai/gpt-oss-20b | 20B | SFT, DPO | LoRA |
| Gemma 3 (27B) VLM | google/gemma-3-27b-it-VLM | 27B | SFT, DPO | LoRA |
| Gemma 4 (31B Instruct) | google/gemma-4-31B-it | 31B | SFT, DPO | LoRA, full |
| Gemma 4 (31B) VLM | google/gemma-4-31B-it-VLM | 31B | SFT, DPO | LoRA |
| Mixtral 8x7B Instruct (v0.1) | mistralai/Mixtral-8x7B-Instruct-v0.1 | 46.7B | SFT, DPO | LoRA, full |
| Llama 3.3 (70B Instruct) | meta-llama/Llama-3.3-70B-Instruct-Reference | 70B | SFT, DPO | LoRA, full |
| Llama 4 Scout (17B-16E Instruct) | meta-llama/Llama-4-Scout-17B-16E-Instruct | 109B | SFT, DPO | LoRA |
| gpt-oss (120B) | openai/gpt-oss-120b | 120B | SFT, DPO | LoRA |
VLM variants accept image columns as multimodal context. Qwen/Qwen3.5-9B (†) is also multimodal despite not carrying the suffix. A full training type means full fine-tuning is available in addition to LoRA.
Per-model training limits
Section titled “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) | Batch size | Gradient accumulation | Maximum lora_r |
|---|---|---|---|---|---|
Qwen/Qwen3.5-0.8B | 131072 | 131072 | 8 | 1 | 64 |
Qwen/Qwen3.5-9B | 65536 | 49152 | 8 | 1 | 64 |
openai/gpt-oss-20b | 131072 | 65536 | 1 | 8 | 64 |
google/gemma-4-31B-it | 49152 | 24576 | 4 | 2 | 64 |
mistralai/Mixtral-8x7B-Instruct-v0.1 | 32768 | 16384 | 8 | 1 | 64 |
meta-llama/Llama-3.3-70B-Instruct-Reference | 24576 | 12288 | 8 | 1 | 64 |
meta-llama/Llama-4-Scout-17B-16E-Instruct | 65536 | 12288 | 8 | 1 | 64 |
openai/gpt-oss-120b | 65536 | 32768 | 2 | 8 | 64 |
Limits for google/gemma-3-4b-it, google/gemma-3-4b-it-VLM, google/gemma-3-27b-it-VLM, and meta-llama/Llama-3.2-3B-Instruct are not published. Leave batch_size unset so AutoScientist resolves the correct value.
Batch size
Section titled “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. This is why gpt-oss trains with a physical batch size of 1–2 and gradient accumulation of 8.
# 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
Section titled “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
Section titled “Row count”AutoScientist requires at least 1,000 rows for every supported model. A smaller model does not lower this platform-wide minimum.
Context length and your data
Section titled “Context length and your data”Training context is the maximum sequence length used during fine-tuning. It is shorter than the serving context and shorter for preference training than supervised fine-tuning.
Rows longer than the training limit are truncated rather than rejected. For datasets containing long documents, choose a model with a larger SFT or DPO context limit.