What is xLLM?
IFM's open-source PyTorch framework for pre-training and fine-tuning dense and MoE LLMs.
Who is xLLM best for?
xLLM is best for AI labs and researchers who pre-train or fine-tune their own dense or MoE LLMs in PyTorch.
Labs pre-training their own LLMs
FSDP, model and context parallelism, custom CUDA kernels and fused blocks reached 52.1% model FLOPs use on Llama3-8B in IFM's test.
Long-context training
Context parallelism and long-sequence attention backends carried K2 Horizon through training stages up to 512K tokens.
Fine-tuning on chat and tool data
The data loader reads chat and tool-call conversations with assistant-only loss masks and packs documents to cut padding.
Researchers extending K2 Horizon
It ships launch scripts for K2 Horizon from 0.9B to 36B, and the native runtime keeps the original training behaviour if you continue pre-training.
How much does xLLM cost?
xLLM is free and open source under the Apache 2.0 license; the only cost is the GPU time you train on.
- Full source on GitHub, commercial use allowed
- Runs on your own NVIDIA GPUs
- No hosted service or paid tier
xLLM is a framework, not a service, so the cost is the GPU time you run it on. IFM's benchmarks used 32 to 128 NVIDIA H200 GPUs; the quick start trains a 0.9B model on a single GPU. Full pricing on github.com. We checked these prices on .
How do readers rate xLLM?
No one has rated xLLM yet. Be the first.
Sign in to rate. One rating per account.
Is xLLM open source?
Yes. xLLM's code is public under the Apache 2.0 licence, so you can read it, change it and run it yourself.
On GitHub
135stars
- Last commit
- 22 Nov
- Repository age
- 3 months
- Latest release
- 24 Oct
- Licence
- Apache 2.0
- Self-hosted
- No
- Repository
- ifm-ai/xllm
Questions about xLLM
What is xLLM?
IFM's open-source PyTorch framework for pre-training and fine-tuning dense and MoE LLMs.
Who is xLLM for?
xLLM is best for AI labs and researchers who pre-train or fine-tune their own dense or MoE LLMs in PyTorch.
Is xLLM free?
Yes. xLLM has a free Open source plan: full source on GitHub, commercial use allowed.
Is xLLM open source?
Yes. xLLM's code is public under the Apache 2.0 licence, so you can read it, change it and run it yourself.
Where can I try xLLM?
On its official website, github.com.
Can xLLM fine-tune models, or only pre-train them?
Both. The xLLM data loader takes chat and tool-call conversations with assistant-only loss masks and best-fit packing for supervised fine-tuning, alongside plain text for pre-training. Its README does not mention LoRA or other adapter methods, so it is aimed at full-model training.
What hardware does xLLM need?
xLLM needs NVIDIA GPUs with CUDA 12.8 or later, PyTorch 2.11 or later, and a CUDA development environment to build its native extensions. The quick start trains a 0.9B model on one GPU; IFM's published benchmarks used 32 to 128 H200 GPUs.







