Axolotl develops open-source software for fine-tuning AI models. Its experimental Nebius Serverless Jobs integration supports single-node training on Nebius GPUs, with checkpoints and final weights exported to object storage.
Axolotl reports validation on one H100 configuration in eu-north1 using NousResearch/Llama-3.2-1B. A 30-step LoRA run, independent export size and hash checks, adapter loading, and training resumption in a fresh Job passed. The resume test restored optimizer, scheduler and random-number-generator state. These results cover one tested configuration. Automatic checkpoint selection and recovery after preemption are not supported.
Commercial use. Commercial classification follows Axolotl's business relationship; its software is open source.
Sources
1 source
Nebius Serverless Jobs validation and deployment guide
Listed 28 Oct 2026 · Axolotl · Web page
Axolotl reports completed validation using Nebius Serverless Jobs on an H100, including LoRA training, exported-file integrity checks, adapter loading and training-state restoration in a fresh Job.
Nebius Serverless Jobs validation and deployment guideListed 28 Oct 2026 · Axolotl · Web pageAxolotl reports completed validation using Nebius Serverless Jobs on an H100, including LoRA training, exported-file integrity checks, adapter loading and training-state restoration in a fresh Job.
Axolotl, developed by Axolotl AI Corp., provides an open-source framework for developers and machine-learning teams to customize language, vision-language and audio models using their own datasets. Its main tool supports full fine-tuning, LoRA and QLoRA, preference learning, reinforcement learning and reward modeling. Users configure data preparation, training, evaluation and inference through YAML files. Axolotl runs on local hardware or cloud infrastructure, with support for training across multiple GPUs and machines, and distributes Python packages and Docker images.
Axolotl v0.21.0 adds Beam fine-tuning and native diffusion language model training
Axolotl released v0.21.0 with QLoRA fine-tuning for Reflection's 501-billion-parameter Beam model on a single eight-H200 node. The release also added native diffusion language model training and image datasets for DPO, KTO and GRPO.
Axolotl v0.20.0 introduces GGUF export and expanded distributed training
Axolotl released v0.20.0 with GGUF checkpoint export, Ringmaster context parallelism, native NVFP4 LoRA and expert parallelism without DeepEP. GGUF export supports deployment through llama.cpp, Ollama, LM Studio and llamafile.
Axolotl v0.19.0 adds new model families and ternary training
Axolotl released v0.19.0 with seven new model families and native ternary quantization-aware training. It also introduced a declarative model-support system and two optimizers.
Open-source software for fine-tuning AI models. · Infrastructure
How they use Nebius AI Cloud
Axolotl develops open-source software for fine-tuning AI models. Its experimental Nebius Serverless Jobs integration supports single-node training on Nebius GPUs, with checkpoints and final weights exported to object storage.
Axolotl reports validation on one H100 configuration in eu-north1 using NousResearch/Llama-3.2-1B. A 30-step LoRA run, independent export size and hash checks, adapter loading, and training resumption in a fresh Job passed. The resume test restored optimizer, scheduler and random-number-generator state. These results cover one tested configuration. Automatic checkpoint selection and recovery after preemption are not supported.
Commercial use. Commercial classification follows Axolotl's business relationship; its software is open source.
Sources
1 source
Nebius Serverless Jobs validation and deployment guide
Listed 28 Oct 2026 · Axolotl · Web page
Axolotl reports completed validation using Nebius Serverless Jobs on an H100, including LoRA training, exported-file integrity checks, adapter loading and training-state restoration in a fresh Job.
Nebius Serverless Jobs validation and deployment guideListed 28 Oct 2026 · Axolotl · Web pageAxolotl reports completed validation using Nebius Serverless Jobs on an H100, including LoRA training, exported-file integrity checks, adapter loading and training-state restoration in a fresh Job.
Axolotl, developed by Axolotl AI Corp., provides an open-source framework for developers and machine-learning teams to customize language, vision-language and audio models using their own datasets. Its main tool supports full fine-tuning, LoRA and QLoRA, preference learning, reinforcement learning and reward modeling. Users configure data preparation, training, evaluation and inference through YAML files. Axolotl runs on local hardware or cloud infrastructure, with support for training across multiple GPUs and machines, and distributes Python packages and Docker images.
Axolotl v0.21.0 adds Beam fine-tuning and native diffusion language model training
Axolotl released v0.21.0 with QLoRA fine-tuning for Reflection's 501-billion-parameter Beam model on a single eight-H200 node. The release also added native diffusion language model training and image datasets for DPO, KTO and GRPO.
Axolotl v0.20.0 introduces GGUF export and expanded distributed training
Axolotl released v0.20.0 with GGUF checkpoint export, Ringmaster context parallelism, native NVFP4 LoRA and expert parallelism without DeepEP. GGUF export supports deployment through llama.cpp, Ollama, LM Studio and llamafile.
Axolotl v0.19.0 adds new model families and ternary training
Axolotl released v0.19.0 with seven new model families and native ternary quantization-aware training. It also introduced a declarative model-support system and two optimizers.