We are excited to share a new milestone — we've open-sourced dInfer, a high-performance inference framework for diffusion language models (dLLMs). 🚀10.7X speedup over NVIDIA’s diffusion model framework Fast-dLLM. 🧠1,011 tokens per second in single-batch inference — on the
2
37
255
Replies
@TheInclusionAI The 1,011 tokens/second on HumanEval is remarkable code generation speed. What's really fascinating is how dInfer manages to outperform even highly optimized autoregressive frameworks in single-batch scenarios - this challenges some conventional wisdom about model architectures
1
0
4
Tiny Recursive Models: A tiny 7M parameter model that recursively refines its answer beats LLMs 100x larger on hard puzzles like ARC-AGI We independently reproduced the paper, corroborated results, and released the weights + API access for those looking to benchmark it 🔍
21
80
584
Playing around with training a tiny 11M parameter character-level text diffusion model! It's a WIP but the code is currently a heavily modified nanochat gpt implementation (to change from autoregressive decoding to diffusion) and trained on the Tiny Shakespeare dataset. The
47
143
2K
🎁🎁 We release Pico-Banana-400K, a large-scale, high-quality image editing dataset distilled from Nana-Banana across 35 editing types. 🔗 Data link: https://t.co/mi06ddf3mN 🔗Paper link: https://t.co/AaZM02xcJr It includes 258K single-turn image editing data, 72K multi-turn
5
57
384
a bunch of OCR models released in past few weeks: ~ deepseek-ocr-3b ~ olmo-ocr-2-7b ~ chandra-ocr-8b ~ nanonets-ocr2-3b ~ paddleocr-vl-0.9B ~ qwen3-vl-dense/moe (general vlm) ~ dots.ocr-3b Will be dropping a detailed comparison soon
34
109
1K