DFloat11 — The Drop-in Successor to bf16 🏆

30 % smaller · 0 % accuracy loss · ~50 % cheaper to serve

Why DFloat11 instead of bf16?

What’s in it for OSS devs & researchers?

• Run bigger models on the same card
  A 24 GB GPU jumps from 8 B → 12 B params without code tweaks.

• Perfect reproducibility
  DF11 is mathematically lossless—identical logits across machines, great for shared baselines and papers.

• No new tooling to learn
  Works out-of-the-box with transformers + gguf; fully LoRA/QLoRA-compatible.

• Open & hackable
  Reference C++/CUDA codec under Apache-2.0 → https://github.com/LeanModels/DFloat11 (PRs welcome!).

• Greener inference
  Less memory traffic = lower power draw—helpful on laptops, edge devices, and big clusters alike.

How does it work?

DFloat11 keeps the sign + mantissa bits unchanged and Huffman-compresses the mostly-empty exponent bits of bf16 weights. Average cost: ≈ 11 bits instead of 16—no information discarded.

Try the first DFloat11 model sets

• FLUX.1
• Gemma 3
• Qwen 3

🔗 Deep-dive blog: “Pied Piper is Here with DFloat11” – diagrams, math, real-world redaction case study.

Lossless • Same transformers workflow • Fine-tune via LoRA/FeatherTune • Open spec

Benchmarks & formal proofs: see the arXiv preprint.

The xMADified Family

From xMAD.ai

Welcome to the official Hugging Face organization for xMADified models from xMAD.ai!

The repositories below contains popular open-source models xMADified with our NeurIPS 2024 methods from 16-bit floats to 4-bit integers, using xMAD.ai proprietary technology.

These models are fine-tunable over the same reduced (4x less) hardware in mere 3-clicks.

Watch our product demo here

CLICK HERE TO JOIN BETA for:

• No-code deployment
• Proprietary Dataset Management
• On-Premise Fine-tuning
• Endpoint Scaling
• System Health Monitoring
• Seamless API Integration

and more!

The memory and hardware requirements (GPU memory needed to run as well as fine-tune them) are listed in the table below:

xMAD.ai LinkedIn