Cortex Training models

Cortex Training supports a curated set of third-party, open models. Use the canonical model ID (for example, Qwen/Qwen3-8B) as the model_name when you create a job. Copy a model ID directly from any card below.

Each card shows whether inference and training are supported. For supported training models, the card summarizes the maximum context length for LoRA (low-rank adaptation) and full-parameter training.

Select View recommended configurations on a card for a copyable Python starting configuration. Select supervised fine-tuning (SFT) or reinforcement learning (RL), then use the Use LoRA checkbox to switch between the complete recommended LoRA and full-parameter profiles when both are available. For models with standard and long-context profiles, use the Enable long sequence checkbox to switch between the model’s two context tiers. The profiles can differ in job topology and distributed training settings, not only in the PEFT configuration. Tune batch sizes and optimization parameters for your workload.

Note

Model availability and specifications are subject to change during the preview.

Available for training

These open weights models support fine-tuning. They also support inference.

Important

Saving and loading LoRA checkpoints for Qwen/Qwen3.6-35B-A3B requires the arctic model provider. LoRA for this mixture-of-experts model currently supports the attention projections (q_proj, k_proj, v_proj, and o_proj). Routed expert LoRA is not yet supported. The Hugging Face MLP names gate_proj, up_proj, and down_proj do not select routed expert parameters in the model implementation used by the arctic provider.

DeepSeek-V4.1-Flash-1017

deepseek-ai/DeepSeek-V4.1-Flash-1017
License: MIT
InferenceSupported
256K max8 GPUs
TrainingSupported
LoRA 256K max · Full 256K max

gemma-3-4b-it

google/gemma-3-4b-it
License: Gemma
InferenceSupported
128K max8 GPUs
TrainingSupported
Full 128K max

gemma-4-31B-it

google/gemma-4-31B-it
License: Gemma
InferenceSupported
256K max8 GPUs
TrainingSupported
Full 256K max

GLM-5.3-Flash

zai-org/GLM-5.3-Flash
License: MIT
InferenceSupported
256K max8 GPUs
TrainingSupported
LoRA 256K max

GLM-5.3-Flash-BF16

zai-org/GLM-5.3-Flash-BF16
License: MIT
InferenceSupported
256K max16 GPUs
TrainingSupported
Full 256K max

Inkling-2-Small-MXFP8

thinkingmachines/Inkling-2-Small-MXFP8
License: Apache 2.0
InferenceSupported
256K max8 GPUs
TrainingSupported
LoRA 256K max

medgemma-1.5-4b-it

google/medgemma-1.5-4b-it
InferenceSupported
128K max8 GPUs
TrainingSupported
Full 128K max

medgemma-4b-it

google/medgemma-4b-it
InferenceSupported
128K max8 GPUs
TrainingSupported
Full 128K max

NVIDIA-Nemotron-3-Nano-30B-A3B-BF16

nvidia/NVIDIA-Nemotron-3-Nano-30B-A3B-BF16
InferenceSupported
256K max8 GPUs
TrainingSupported
Full 256K max

NVIDIA-Nemotron-3.5-Lightning-30B-A3B-BF16

nvidia/NVIDIA-Nemotron-3.5-Lightning-30B-A3B-BF16
License: OpenMDW-1.1
InferenceSupported
256K max8 GPUs
TrainingSupported
Full 256K max

Qwen3-0.6B

Qwen/Qwen3-0.6B
License: Apache 2.0
InferenceSupported
32K max1 GPU
TrainingSupported
LoRA 32K max · Full 32K max

Qwen3-1.7B

Qwen/Qwen3-1.7B
License: Apache 2.0
InferenceSupported
32K max8 GPUs
TrainingSupported
LoRA 32K max · Full 32K max

Qwen3-30B-A3B

Qwen/Qwen3-30B-A3B
License: Apache 2.0
InferenceSupported
40K max8 GPUs
TrainingSupported
Full 40K max

Qwen3-4B-Instruct-2507

Qwen/Qwen3-4B-Instruct-2507
License: Apache 2.0
InferenceSupported
256K max8 GPUs
TrainingSupported
Full 256K max

Qwen3-8B

Qwen/Qwen3-8B
License: Apache 2.0
InferenceSupported
32K max2 GPUs
TrainingSupported
LoRA 32K max · Full 32K max

Qwen3.5-2B

Qwen/Qwen3.5-2B
License: Apache 2.0
InferenceSupported
256K max8 GPUs
TrainingSupported
LoRA 256K max · Full 256K max

Qwen3.5-4B

Qwen/Qwen3.5-4B
License: Apache 2.0
InferenceSupported
256K max8 GPUs
TrainingSupported
LoRA 256K max · Full 256K max

Qwen3.5-9B

Qwen/Qwen3.5-9B
License: Apache 2.0
InferenceSupported
256K max8 GPUs
TrainingSupported
LoRA 64K max · Full 256K max

Qwen3.6-35B-A3B

Qwen/Qwen3.6-35B-A3B
License: Apache 2.0
InferenceSupported
256K max8 GPUs
TrainingSupported
LoRA 256K max · Full 256K max

Qwen3.8-27B

Qwen/Qwen3.8-27B
License: Apache 2.0
InferenceSupported
256K max8 GPUs
TrainingSupported
LoRA 256K max · Full 256K max

Inference only

These models support inference (sampling) but not fine-tuning.

GLM-5.3

zai-org/GLM-5.3
InferenceSupported
704K max8 GPUs
TrainingNot supported

gpt-oss-120b

openai/gpt-oss-120b
License: Apache 2.0
InferenceSupported
128K max8 GPUs
TrainingNot supported