transformer

huggingface/transformers

A unified model-definition framework spanning text, vision and audio — Transformers acts as the shared definition layer for models like DeepSeek, Gemma, GLM and Qwen, supporting both inference and training across text, vision, audio and multimodal tasks. It suits researchers and engineers who want one consistent API and access to pretrained checkpoints from the Hub instead of per-architecture code.

vllm-project/vllm

vLLM keeps growing as a broad, hardware-agnostic LLM serving engine — vLLM is a high-throughput inference engine using PagedAttention, continuous batching, and quantization (FP8, INT4, GGUF, and more) to serve LLMs cheaply. Backed by over 2000 contributors, it now supports CUDA, AMD, TPU and Blackwell hardware plus models like DeepSeek-V3, Qwen3, and GPT-OSS, making it suited for teams deploying large models at scale.

labmlai/annotated_deep_learning_paper_implementations

A literate-programming library of 60+ annotated deep learning papers — This repo pairs PyTorch code with side-by-side explanatory notes for transformers, GANs, RL algorithms, optimizers and more, rendered on a companion website. It's actively updated weekly and now points to a separate 'Annotated Implementation Skill' for using coding agents to implement new papers in the same style. Useful for learners and engineers wanting to read algorithms as documented code rather than just papers.

p-e-w/heretic

Automatic abliteration tunes away LLM refusals, no fine-tuning — Heretic automates directional ablation with an Optuna TPE optimizer to strip safety alignment from transformer LLMs, balancing refusal rate against KL divergence to preserve model quality. No ML expertise needed—just a CLI—and it covers dense, MoE, multimodal, and hybrid architectures like Qwen3.5.

sgl-project/sglang

Unified inference engine for LLMs, VLMs, and diffusion models — SGLang now bundles SGLang Diffusion, a built-in image/video generation engine, alongside its existing LLM and vision-language serving stack. It targets agentic workloads, large-scale serving, and RL rollouts, with broad hardware and model support including Blackwell GPUs, MoE architectures, and models like DeepSeek, Qwen, and GLM.

fishaudio/fish-speech

SOTA Open Source TTS

SYSTRAN/faster-whisper

Faster Whisper transcription with CTranslate2

FlashML-org/FreeToken

Serving 290B-parameter MoE models on a gaming PC — FreeToken is an edge-native inference engine that lets large Mixture-of-Experts models like DeepSeek-V4, GLM, Minimax and Qwen run on consumer GPUs/CPUs instead of datacenters. It uses bandwidth-adaptive CPU-GPU co-execution, expert caching, and elastic VRAM management to keep 290B+ models interactive on a single machine.

FareedKhan-dev/kimi-k3-in-c

2.78T-parameter Kimi K3 inference in 176 KB of C, no GPU needed — This project runs the 2.78-trillion-parameter Kimi K3 model on a single CPU with only 8.24 GB peak RAM, using a 176 KB portable C99 engine with no BLAS or framework dependency. It streams the 1.56 TB checkpoint from disk, so more memory just speeds up inference rather than being required to run it at all. Useful for anyone curious about running huge MoE/transformer models on ordinary hardware.

vllm-project/semantic-router

A programmable routing layer that picks the right LLM per call — vLLM Semantic Router gives agent harnesses one API across many models and backends, choosing or combining them per request based on policy for quality, latency, cost, and location. Useful for teams running mixed local, private, and cloud inference who want a single decision layer instead of hardcoding model choices.

Interactive-Active-AI/NextTransformer

:sunrise: A comprehensive, cutting-edge open-source framework for next-generation Transformers/LLM, with compute-matched benchmarking and a platform for autonomous architecture research.

mottopanikeiku/doomformer

A Neural Network Compiled to Run Inside DOOM Itself — Doomformer compiles integer neural networks into ZDoom maps, using voodoo doll arithmetic and sector light levels/heights to store values, then runs them live in ViZDoom. Every computed value is read back and checked against a NumPy reference with zero tolerance. It can classify a handwritten digit, though up to 65,000x slower than NumPy—an esoteric computing experiment for curious engineers.