Salvatore Sanfilippo

Author of Redis

Authored Projects (7)

Starred by

Balaji Srinivasan(Founder of The Network School; Author of "The Network State"; Former CTO of Coinbase; Cofounder of Counsyl),

Jason Miller(Author of Preact),

Abubakar Abid(Cofounder of Gradio),

Jeff Hammerbacher(Cofounder of Cloudera), and

13 more.

ds4 by antirez

0.0%

19k

Fast local inference for DeepSeek V4 Flash models

Enables high-performance local inference of large language models on Apple Silicon using Metal.
Supports massive 1M token context windows with compressed KV caches persisted to disk.
Integrates with AI agents via OpenAI/Anthropic compatible APIs, supporting advanced quantization.
Features AI-assisted development and a Metal-only focus, with CPU path for debugging only.

Created 3 months ago

Updated 3 weeks ago

Starred by

Quinn Slack(Cofounder of Sourcegraph)

Georgi Gerganov(Author of llama.cpp, whisper.cpp)

Simon Willison(Coauthor of Django)

gguf-tools by antirez

350

CLI tool for GGUF file manipulation and analysis

Shows detailed info about GGUF files, key-value pairs, and tensors.
Compares two GGUF files, computes average weights difference (%).
Inspects tensor weights, useful for quantization checks/fingerprinting.
Splits Mixtral MoE models (experimental), extracts 7B model subset.

Created 2 years ago

Updated 3 months ago

llama.cpp-deepseek-v4-flash by antirez

329

Experimental LLM inference engine for efficient local deployment

Leverages C/C++ and advanced quantization for high-performance, low-memory LLM inference.
Supports multimodal models and numerous text LLMs, featuring experimental DeepSeek v4 Flash.
Offers an OpenAI-compatible API server, CLI tools, and bindings for seamless application integration.
Optimized for resource-constrained environments, enabling LLMs on devices with limited RAM.

Created 4 months ago

Updated 4 months ago

qwen-asr by antirez

582

Pure C ASR inference engine

Fast speech-to-text for Qwen3 models, supporting offline and streaming modes.
Optimized C implementation with minimal dependencies, leveraging BLAS for CPU performance.
Near-instant model loading via memory-mapped safetensors and efficient streaming.
Handles piped audio input, offering token-by-token output and language control.

Created 6 months ago

Updated 6 months ago

Starred by

Dan Guido(Cofounder of Trail of Bits),

Patrick von Platen(Author of Hugging Face Diffusers; Research Engineer at Mistral),

Luis Capelo(Cofounder of Lightning AI),

Phil Wang(Prolific Research Paper Implementer), and

1 more.

voxtral.c by antirez

Pure C speech-to-text inference engine for Mistral Voxtral Realtime 4B

Enables real-time, low-latency speech transcription with a streaming C API.
Leverages Metal GPU acceleration on Apple Silicon or BLAS for efficient inference.
Features memory-mapped weights and a rolling KV cache for fast loading and unlimited audio length.
Offers a zero-dependency C implementation, runnable without Python, CUDA, or vLLM.

Created 6 months ago

Updated 6 months ago

Starred by

Luis Capelo(Cofounder of Lightning AI),

Simon Willison(Coauthor of Django),

Michael Chiang(Cofounder of Ollama),

Theo Browne(Founder of Ping.gg), and

2 more.

iris.c by antirez

Pure C inference engine for AI image generation

Generates images from text prompts and transforms existing images using the FLUX.2-klein-4B model.
Supports text-to-image and image-to-image modalities with an integrated Qwen3-4B text encoder.
Zero-dependency C inference avoids Python/PyTorch/CUDA, with optional BLAS/MPS GPU acceleration.
Features memory-efficient operation and a low-memory mode (--mmap) for systems with limited RAM.

Created 7 months ago

Updated 6 months ago

Starred by

Boris Cherny(Creator of Claude Code; MTS at Anthropic),

Luis Capelo(Cofounder of Lightning AI),

Tim Suchanek(Founder of expand.ai),

Simon Willison(Coauthor of Django), and

5 more.

neural-redis by antirez

Redis module for feed-forward neural networks as a native data type

Simplifies ML workflows by integrating data collection and training within Redis.
Uses resilient backpropagation (RPROP) for online training in separate threads.
Supports automatic data normalization and overtraining detection.
Enables regression and classification tasks, suitable for web/mobile app enhancements.

Created 10 years ago

Updated 8 years ago