Open-source deep learning,
built in public.
Compact models. Efficient architectures. Open evaluation tools.
An R&D collective — not a company, not incorporated.
What we research
Liodon is an independent, self-funded applied research lab working on the frontier of open-source foundation model research.
Models & Architecture
Small Language Models
Architectural efficiency via Transformer variants, sparse attention, quantization, and knowledge distillation — targeting models that run on constrained hardware.
- SLM-10M — tops Open SLM Leaderboard <10M tier
- whisper-large-v3-te — Telugu ASR fine-tune
- GGUF variants of Gemma 4 & Qwen
- slm-bench evaluation dataset
Python Libraries
Research Tools & Libraries
Purpose-built Python packages for the deep learning researcher — from loss functions and LLM evaluation to agentic primitives and citation auditing.
- JuryEval — LLM-as-Judge evaluation framework
- TorchCriterion — modular loss functions
- TorchAgentic — agentic AI primitives
- CitationHunter, PromptGrad, NanoNLP & more
Triton Kernels · Featured
TorchEmbed
Fused Triton kernels for RoPE and embedding operations — 4–7× faster than pure PyTorch and ~2× faster than torch.compile. Used by DeepSpeed and Microsoft.
- RoPE fused forward & backward kernels
- 4–7× faster than pure PyTorch
- ~2× faster than torch.compile
- Used by DeepSpeed & Microsoft
Triton-fused performance
Measured on NVIDIA GB10 (float16). The fused kernel reads and writes each element exactly once — no intermediate allocations.
82+
Models, Datasets & Repos
100%
Open Source (MIT / Apache 2.0)
46,000+
Downloads on HuggingFace
TorchEmbed · used by
Featured Projects
Practical tools for the research community — all open source, all on GitHub.
Language Model
SLM-10M
10M parameter SLM leading the Open SLM Leaderboard under 10M tier. Efficient, quantizable, and freely available on HuggingFace.
Evaluation Library
JuryEval
LLM-as-Judge evaluation with statistical significance testing. Go beyond single-judge prompts with multi-judge consensus.
Research Tool
CitationHunter
Audits research papers for missing or weak citations. Built for researchers who need thorough literature grounding.
From the blog
Research notes, model releases, and open-source updates.
- TorchEmbed in DeepSpeed: Fused RoPE Kernels at ScaleTorchEmbed is now integrated into DeepSpeed and ships as a dependency. The fused Triton RoPE kernel replaces DeepSpeed’s transformer engine’s rotary embedding path, delivering 4–7× speedup over pure PyTorch with no code changes required from users.
- Fine-tuning Whisper Large v3 for Telugu ASR: whisper-large-v3-teTelugu is spoken by roughly 80 million people, but remains severely under-resourced in the ASR landscape. whisper-large-v3-te is our fine-tuned version of OpenAI’s Whisper Large v3, trained specifically for Telugu… Read more: Fine-tuning Whisper Large v3 for Telugu ASR: whisper-large-v3-te
- JuryEval: LLM-as-Judge Evaluation with Statistical Significance TestingSingle-judge LLM evaluation is noisy. Ask GPT-4 to rate two outputs and the ordering can flip on a re-run. JuryEval addresses this by running multiple LLM judges in parallel, then… Read more: JuryEval: LLM-as-Judge Evaluation with Statistical Significance Testing
Ready to explore the research?
Liodon AI is not a company — just open-source research built in the open. Star a repo, open an issue, or just read the code.