CatalogWantedMirrorsLedgerOperating StatementDocsStudio statusCard system

Open models stay open.

Find, check and download open AI models. Free. No account. By Allternit, the open source alternative.

Browse the catalogRead the Operating Statement
$ HF_ENDPOINT=https://freeweights.allternit.com hf download Qwen/Qwen3-Embedding-0.6B
83
Models listed
Listed means downloadable. Read from the catalog.
2,412
Files with a signed sha256
Every file in a signed manifest.
Not live
Download starts
Counted by the download Worker from launch day. No number until then.

How it works

1. Find

Search by name, task, license, format or size. Models are grouped by family. Order is never paid.

2. Check

Every model has a license and a signed manifest. Every file has a size and a sha256. You can check them with stock tools.

3. Download

Use our fw.py tool, the Hugging Face tools with one setting, or plain links. No account, no token.

4. Know the source

A badge says where the bytes come from: Hosted by Allternit, Community seeded, or From source. We check it when you ask, and a model with no live source moves to Wanted.

Collections

Full catalog →

Worked out from the real files in each signed manifest, not from model names.

Runs on a 256 GB MacWeights of 200 GB or less, counted from each signed manifest (the smallest GGUF or ONNX variant when a repo has several), so the model fits in 256 GB of unified memory with room for context.60 modelsSmall enough for a laptopWeights of 12 GB or less, counted from each signed manifest, so the model loads in 16 GB of memory with room to run.27 modelsEmbeddings and rerankersModels that turn text or images into vectors, and models that rank search results.9 modelsSpeechSpeech recognition, text to speech and audio models.8 modelsGGUF for llama.cpp and OllamaRepos with GGUF files, ready for llama.cpp, Ollama and LM Studio.11 modelsOpen datasetsTraining and evaluation datasets with a license that allows sharing.9 models

Latest additions

Newest first →
From: ACE-Step <huggingface.co/ACE-Step>Subject: Ace-Step1.5
From: unsloth <huggingface.co/unsloth>Subject: DeepSeek-V4.1-Flash-1017-GGUF
From: black-forest-labs <huggingface.co/black-forest-labs>Subject: flux-3-image-klein-comfy
From: alesha-pro <huggingface.co/alesha-pro>Subject: GLM-5.3-Flash-abliterated-Maya-M-GGUF
From: huihui-ai <huggingface.co/huihui-ai>Subject: Huihui-DeepSeek-V4.1-Flash-1017-abliterated-GGUF
From: huihui-ai <huggingface.co/huihui-ai>Subject: Huihui-Qwen3.8-27B-abliterated-GGUF

Sections

All 22 sections →

The catalog is grouped into sections. Each section is a family of models, newest first. Modality is a filter, not a section.

alt.freeweights.qwen

Qwen language, vision, embedding, speech and image models.

23 models: Qwen3.8-Flash-Next-GGUF, MiMo-V2.6-Distill-Qwen-9Balt.freeweights.tools

Inference and training tools, linked to their source repos.

12 tools: ComfyUI, llama.cppalt.freeweights.datasets

Training and evaluation datasets with open licenses.

9 models: fineweb-edu, gpqaalt.freeweights.gemma

Google Gemma models.

6 models: embeddinggemma-2, gemma-4-12B-italt.freeweights.deepseek

DeepSeek models and community GGUF builds.

5 models: DeepSeek-V4.1-Flash-1017-GGUF, DeepSeek-V4-Flash-0731alt.freeweights.speech

Speech recognition and text-to-speech models.

5 models: s2-pro, chatterbox

Check every byte

A model is three things you can check with stock tools: a signed manifest, the sha256 of each file, and the license text.

  1. The manifest is signed. minisign, Ed25519, key 2DC28CC7B26A283B.
  2. Every file has a size and sha256. A file that does not match is thrown away and the next source is tried.
  3. The upstream commit is pinned. The bytes are the publisher's release at that commit.
curl -O https://freeweights.allternit.com/manifests/ACE-Step/Ace-Step1.5.json
curl -O https://freeweights.allternit.com/manifests/ACE-Step/Ace-Step1.5.json.minisig
minisign -Vm Ace-Step1.5.json -P RWQ7KGqyx4zCLXtUFOqvfv8L3g+8fpKJ1FsuxPyOYMF+4KufbUxKq1Oq
How to verify a download

Why you can trust it

  • We publish the license and a checksum with every file.
  • The manifests are signed. You verify with a public key.
  • The code is open and the costs are public. The ledger is a signed hash chain.
  • We never host sexual-content models, child sexual abuse material, real-person face or voice models without documented consent, or models marketed for fraud, malware or attacks.
  • Support goes first to running costs. It is not a charitable gift and is not tax-deductible.

Read the full Operating Statement for the promises and how we keep them.

Ledger0 USDCosts, 2026-11. 6 entries in the signed hash chain, head 6c8154c7…e0b6. Goal 10 USD of 45 USDStage 1: Tracker and hub. Support is not open yet. Mirrors0No mirrors yet. Have spare disk and bandwidth? Run the first one.

The operating statement

Read the operating statement →

Freeweights keeps open AI models open. Anyone can find, check and download them for free, with no account. Allternit, an American company, runs it in the open. The code is open source and the costs are public. Freeweights is subsidized: supporters, sponsors, mirror hosts and a share of Allternit's revenue pay for it.