Unsloth
App in the BluixApps catalog
What it is
Unsloth is the fastest LLM fine-tuning library — custom Triton kernels deliver 2× the speed and 50% less VRAM than vanilla HuggingFace + PEFT. Maintained by Unsloth AI (Daniel Han, ex-Microsoft). The library of choice when budget GPU + speed matter.
For solo developers and AI tinkerers fine-tuning on Colab/consumer GPUs, Unsloth is the canonical choice.
What it's for
- Lightning-fast LoRA training — 2-5× faster than alternatives
- Low VRAM training — 7B QLoRA on 8 GB VRAM (vs 24 GB elsewhere)
- Pre-quantized models — load 4-bit base instantly (no quantize-at-load delay)
- Native multi-GPU — added Q4 2024
- Broad model support — Llama, Mistral, Qwen, Phi, Gemma all covered
- TRL integration — SFT, DPO, ORPO via TRL trainers
Who it's for
- Solo AI developers fine-tuning on consumer GPUs
- Researchers running fine-tuning experiments on a budget
- Startups wanting fastest iteration on training experiments
- Educators running fine-tuning workshops on shared hardware
- Hosting providers offering low-cost fine-tuning tier
Why teams pick Unsloth over alternatives
- Apache 2.0 — fully open
- Fastest — 2-5× speedup vs standard transformers + PEFT
- Lowest VRAM — 50% less than alternatives
- Pre-quantized HF models at
unsloth/*namespace (instant load) - Active development — frequent releases, Triton kernel optimizations
- Daniel Han backing — known LLM optimization expert
- Notebook library — Colab-ready examples for common tasks
Integrations
- HuggingFace Transformers — base
- PEFT + TRL — LoRA + SFTTrainer
- Pre-quantized models at
unsloth/*HF namespace - Pair with: vLLM/TGI to serve fine-tuned (Unsloth → save_pretrained_merged → load with vLLM)
- DPO/ORPO support via TRL
- Continued pretraining for domain-adapt
Notable users & community
- 24k+ GitHub stars
- Unsloth AI corporate backing
- Daniel Han (ex-Microsoft) leads development
- Featured in popular Colab fine-tuning tutorials
- Active Discord + Reddit presence
Tips & operations
- VRAM with Unsloth:
- 7B QLoRA: 8 GB VRAM minimum (!)
- 13B QLoRA: 12 GB
- 70B QLoRA: 48 GB (vs 80 GB standard)
- Pre-quantized models:
unsloth/Meta-Llama-3.1-8B-Instruct-bnb-4bitloads in seconds - Code pattern: from
FastLanguageModel.from_pretrained()→get_peft_model()→ TRLSFTTrainer - Multi-GPU: enable in newer versions via
tensor_parallel - Save:
model.save_pretrained_merged()to export combined weights - vs Axolotl: Unsloth = code/library, Axolotl = config-driven. Use Unsloth for speed-critical custom code; Axolotl for reproducible config workflows
- vs LLaMA-Factory: Unsloth = library; LLaMA-Factory = visual UI on top
What we ship in BluixApps
- Docker (pytorch base + Unsloth pip-installed at runtime)
- JupyterLab pre-installed for interactive notebooks
- Persistent volumes: workspace, datasets, outputs
- Port 8889 mapped
- Pre-set HF_TOKEN environment variable for gated models
- Install report at
/root/bluixapps/unsloth.txt - Full Python quick-start example (paste into Jupyter)
- Notebook library URL for premade Colab-ready examples
- Pairing notes (vLLM/TGI for serving merged model)
- GPU pre-flight check via
bluixapps_ensure_nvidia_runtime - Backup hook covers workspace + outputs
Get this app — pick a BluixApps plan
Same catalog. Scaling tenant isolation, white-label and support tier.
| Tier | Tenants | Catalog | Support | White-label | Monthly | |
|---|---|---|---|---|---|---|
| Stacks | 1 | 19 curated stacks | Standard | — | $19/mo | DetailDeploy |
| Starter | 10 | Full catalog | Standard | +$15–25/mo | $49/mo | DetailDeploy |
| Pro | 25 | Full catalog | Priority bugfix | +$15–25/mo | $149/mo | DetailDeploy |
| Growth | 100 | Full catalog | Priority bugfix | +$15–25/mo | $349/mo | DetailDeploy |
| Scale | 500 | Full catalog | 7-day window | +$15–25/mo | $799/mo | DetailDeploy |
| Enterprise | Unlimited | Full catalog | Priority 7-day | Bundled | $1,499/mo | DetailDeploy |
