open-weights

Easily fine-tune, evaluate and deploy Qwen, Gemma, or any open weight LLM!

MODA: open fashion retrieval benchmark and models by Hopit AI. MODA (203M, open source), MODA Pro Lite (213M, open weights), MODA Pro (hosted). Full-corpus benchmarks vs FashionSigLIP, SigLIP-SO400M and ZooClaw — one harness, losses shown. #1 open model on LookBench.

Curated Ideogram 4.0 prompts and image examples — typography, photorealistic portraits, product/UI mockups, and model comparisons. By Evolink.

WicaraLLM is our Indonesian Small Language Model (56M parameters), built from scratch and trained on a 1.3B-token Indonesian corpus. Engineered to run efficiently on a single laptop GPU with just 6 GB of VRAM, it combines GQA, RoPE, SwiGLU, and RMSNorm, with supervised fine-tuning for conversational capabilities

Introducing Muse Glimmer: open-weight 30B agentic multimodal model that runs on your device (Meta). Interactive local agent lab + guide. Apache 2.0 · on-device AI · function calling

One embedding space for text, image, video, audio, thermal, motion, and touch. Open weights, self-hostable.

OpenMayhem is an open-source and decentralized inference router for consumer & pro hardware. Providers earn, users/agents access cheap inference.

Companion code for the Manning book 'LLM Customization and Fine-Tuning.' Adapt open-weights LLMs end to end: prompting and RAG, LoRA/QLoRA, full SFT, distillation, and DPO/RLHF alignment.

Open-weight only, video-aware podcast dubbing with ASR alignment, diarization, LLM translation, and voice-cloned TTS. Local inference.

A real engineering language model running fully offline on an $8-class ESP32-S3 — open weights, native C, real-silicon benchmarks.

A community-maintained catalog of open, low-filter, and community-reported AI video models, checkpoints, adapters, and endpoints.

An independent, evidence-graded read on how much you own the open models and inference providers you rely on.

Structured YAML catalog of 4,587 AI models across 95 providers — pricing, context windows, modalities, capabilities. First-party data with TypeScript types and Zod validation.

See when Claude Code reroutes your request to a different model — and choose the fallback yourself. 123-line local proxy, zero deps.

Redrob Image: open-weight text-to-image model by Redrob. Free to use, free for commercial use (Apache-2.0).

FastMCP fleet MCP server for diffusion LMs (dLLM). DiffusionGemma on Goliath RTX 4090 — batch inference, HLE-shaped reasoning, ~200–400 tok/s. Doc phase; llama-diffusion-cli sidecar next. Complements local-llm-mcp.

Frontier-Grade Open Weights ── フロンティア級のオープンウェイトモデルは、開かれたのか。Kimi K3 / GLM-5.3 を継続追跡する定点観測型OSS書籍 / They matched the frontier. But no one can hold them. Continuously updated.

Streaming, steerable text 2 video: a Wan2.1-1.3B student distilled with SF DMD that generates video continuously instead of as a fixed clip. (like wanstreamer)

tiny tool, big videos. topic in, video out. no film degree required. webapp, windows app and mcp server for agent use

MiniMax H3 Dropped: FREE Open Source 2K Video AI That BEATS Paid Tools? Lightweight open-weights 2K multimodal video generation workspace with REST API integration.