Which fine-tuning framework is best in 2026?

Unsloth for the fastest single-GPU training with the lowest VRAM. Axolotl for the most flexible YAML-config-driven training. LLaMA-Factory for the broadest feature set and best web UI.

How much VRAM do I need?

For QLoRA 4-bit on a 7B model: 12GB minimum, 16GB comfortable. For full fine-tuning of a 7B: 48GB+. For QLoRA on a 70B: 48GB. Unsloth claims 40% less VRAM than baseline, which often means the difference between fitting and not fitting.

Can I fine-tune a 70B model on a single 24GB GPU?

With QLoRA + Unsloth + 4-bit quantization, a 70B model can be fine-tuned on a single 24GB GPU at very low batch size and gradient accumulation. Speed is slow (~1 step/sec) but it works.

Do these work on Apple Silicon?

Unsloth and Axolotl have experimental Metal support. LLaMA-Factory works on MPS but with limitations. For production Apple Silicon fine-tuning, consider MLX + custom training loop.

Unsloth vs Axolotl vs LLaMA-Factory in 2026

These three frameworks cover the bulk of local LLM fine-tuning in 2026. They differ in speed, VRAM efficiency, flexibility, and feature set. This page helps you pick.

Overview

	Unsloth	Axolotl	LLaMA-Factory
First release	2023	2023	2023
Language	Python (CUDA kernels)	Python (PyTorch + DeepSpeed)	Python (PyTorch + Transformers)
Primary use	single-GPU QLoRA	multi-GPU production	web UI + broad features
Training methods	LoRA, QLoRA, full FT	LoRA, QLoRA, full FT, DPO, RLHF	LoRA, QLoRA, full FT, DPO, PPO, KTO, ORPO, +
VRAM savings (claimed)	40% vs baseline	varies	varies
Speedup (claimed)	2-5×	1-1.5×	1-1.2×
YAML config	✗ (Python)	✓ (extensive)	✓ (extensive)
Web UI	✗ (notebooks)	✗	✓ (LLaMA Board)
Multi-GPU	partial	✓ (DeepSpeed, FSDP)	✓ (DeepSpeed, FSDP)
License	Apache 2.0	Apache 2.0	Apache 2.0

VRAM and speed (7B model, QLoRA, batch 1, 1024 ctx, RTX 4090)

Framework	VRAM used	Tokens/sec
Baseline (HF + PEFT)	18 GB	1,200
Unsloth	11 GB	3,400
Axolotl	16 GB	1,500
LLaMA-Factory	17 GB	1,400

Unsloth has a clear lead on single-GPU. The gap closes on multi-GPU setups where DeepSpeed and FSDP matter more.

When to use each

Choose Unsloth when…

You are fine-tuning on a single GPU (24GB or less).
You want the fastest training and the lowest VRAM.
You are doing LoRA or QLoRA (not full fine-tuning).
You are okay with a Python notebook-style API rather than YAML.

Choose Axolotl when…

You need the most flexible YAML config for complex training setups.
You are doing multi-GPU training with DeepSpeed or FSDP.
You want every training method (DPO, RLHF, ORPO, KTO, etc.) supported.
You are training a research model where the config is the artifact.

Choose LLaMA-Factory when…

You want a web UI to manage experiments.
You want the broadest feature set including the latest alignment methods.
You are doing continued pre-training as well as fine-tuning.
You want a single tool for training, evaluation, and export.

Code comparison

Unsloth (Python notebook)

from unsloth import FastLanguageModel
from trl import SFTTrainer

model, tokenizer = FastLanguageModel.from_pretrained(
    "unsloth/llama-3.1-8b-bnb-4bit",
    max_seq_length=2048,
    load_in_4bit=True,
)

model = FastLanguageModel.get_peft_model(
    model, r=16, target_modules=["q_proj", "k_proj", "v_proj", "o_proj"],
)

trainer = SFTTrainer(model=model, train_dataset=dataset, ...)
trainer.train()

Axolotl (YAML)

# axolotl config
base_model: meta-llama/Llama-3.1-8B-Instruct
model_type: LlamaForCausalLM
load_in_4bit: true
adapter: qlora
lora_r: 16
lora_target_modules:
  - q_proj
  - k_proj
  - v_proj
  - o_proj
datasets:
  - path: my_dataset.jsonl
    type: alpaca

accelerate launch -m axolotl.cli.train config.yml

LLaMA-Factory (YAML or Web UI)

# Web UI
llamafactory-cli webui

# CLI
llamafactory-cli train \
  --model_name_or_path meta-llama/Llama-3.1-8B-Instruct \
  --quantization_bit 4 \
  --lora_rank 16 \
  --dataset my_dataset \
  --output_dir ./output

Decision matrix

Your situation	Use
Single GPU, 12-24GB VRAM	Unsloth
Single GPU, want max speed	Unsloth
Multi-GPU, DeepSpeed / FSDP	Axolotl
Research, complex YAML	Axolotl
Want web UI	LLaMA-Factory
DPO, ORPO, KTO, RLHF	LLaMA-Factory (broadest)
Continued pre-training	LLaMA-Factory
Just need to fine-tune a 7B fast	Unsloth
Production-grade training	Axolotl or LLaMA-Factory
On Apple Silicon	LLaMA-Factory (best MPS support)

Unsloth vs Axolotl vs LLaMA-Factory: Fine-Tuning Frameworks in 2026