Language Models

To prove that a 2.78-trillion-parameter model can run on a single CPU with 8 GB of RAM and no GPU.

WASTE exists to find out how far local inference can be pushed when model weights live mostly on fast storage instead of RAM.

TurboFieldfare streams individual experts from SSD on demand so Gemma 4 26B-A4B fits inside ~2 GB of RAM on any Apple Silicon Mac.

The project squeezes a 28.9-million-parameter language model onto an ESP32-S3 microcontroller by keeping most of its weights in slow flash memory and reading only what each token needs.

Because standard Whisper sanitizes speech, and verbatim use cases actually need every filler, stutter, and false start.

A privacy-first Android fork that runs LLMs, image generation, and speech AI entirely offline, then locks itself behind your fingerprint.

It breaks complex tasks across a team of specialized LLM agents that refine their own skills as they work.

A curated reading list that maps how the post-training world is moving from static SFT to student-rollout distillation with live teacher feedback.

An open-source AI assistant built in Kotlin that remembers context across chats, draws interactive screens, and runs a sandboxed Linux distro on your phone.

A Chinese-language LLM curriculum that rebuilds Stanford CS336 into six hands-on assignments, from writing a tokenizer to distributed training and GRPO.

Agnes AI is a hosted multimodal API that speaks OpenAI's protocol, letting you reroute existing clients to its text, image, video, and agent models by changing a base URL.

TinyEngram open-sources experiments showing that DeepSeek's Engram architecture can inject domain knowledge into Qwen and Stable Diffusion more efficiently than LoRA, with less catastrophic forgetting.

A modular speech-to-speech pipeline that exposes an OpenAI Realtime-compatible WebSocket API so you can run voice agents on local or open-source models instead of proprietary cloud services.

Datus is an open-source agent that keeps LLMs from hallucinating SQL by building a living, learning knowledge base around your data stack.

Nativ bundles an embedded mlx-vlm server into a SwiftUI app to turn your Apple Silicon Mac into a local AI workspace with OpenAI-compatible APIs.

HRM-Text claims to cut pretraining costs by 130–600× compute and 150–900× data, shipping a full 1B-parameter framework with FSDP2, FlashAttention 3, and a hierarchical recurrent architecture.

It split off from `verl` to give diffusion, video, and omni-modality models an RL post-training framework that doesn't treat them like chatbots.

It turns ChatGPT’s browser-only image generation into a poolable, OpenAI-compatible API so you can self-host programmatic access to GPT-Image-2 and friends.

It unifies Stable Diffusion, GGUF chat, Whisper, and Kokoro TTS into a single offline desktop GUI so you can skip cloud APIs, subscriptions, and censorship filters.

LinguaGacha uses LLMs to batch-translate novels, subtitles, and game scripts while auto-generating glossaries so character names stay consistent across an entire work.