What we are building in the open. Every project is released as open source, with the training recipes and eval results alongside, so anyone can check our claims.
Released: v0.2.2
tunekeep
Keep a fine-tuned model working when its provider retires the base model: record it, test a replacement, retrain on open weights if needed.
Hosted fine-tunes are tied to base models that get retired, and the training is lost with them. OpenAI retires its fine-tuned GPT-4o and GPT-4.1 models on 14 January 2027. tunekeep lists every fine-tune you own with its days left, records how each one behaves while it still answers, checks whether a newer model already matches it, and retrains an open model with LoRA from your own data when none does.
Inventory of every fine-tuned model, with shutdown dates
A resumable recording of real answers, kept as an eval set
Scoring of any replacement model against those answers
LoRA retraining of open models on NVIDIA GPUs, exported to GGUF (Qwen3-1.7B: 50% to 91% on banking77)
One self-contained report on quality, speed and tokens
A second opinion for AI photo descriptions: several local vision models describe a photo, and you hear where they agree and where they disagree.
Blind and low-vision people use AI descriptions for labels, medicine and mail, but a single description sounds equally sure when it is wrong. Research at ASSETS 2025 found that showing where models disagree helped blind participants spot 4.9 times as many unreliable claims. Crosscheck brings that to models that run on your own computer, so the photo never leaves it.
Several vision models from different families describe each photo as short facts
Every model checks the others' facts; disagreements on numbers and text are kept apart
A spoken summary first, details on demand, fully keyboard and screen reader friendly
Runs locally through Ollama, with no cloud and no account
Rerank for Ollama: adds the rerank endpoints apps expect, using the reranker models you already pulled, and passes everything else through.
Reranking is what makes search and RAG answers good, and it is the most requested missing piece in Ollama. rerankd sits next to Ollama, answers /api/rerank and /v1/rerank with llama.cpp, and checks every model before loading it. Many shared Qwen3 reranker files are broken, and it says so instead of returning meaningless scores.
Works with Dify, Cherry Studio and anything that speaks the Jina or Cohere rerank API
Rankings match the original Hugging Face models (top result 8 of 8 in our integration test)
Starts the reranker on demand and stops it when idle, on NVIDIA GPUs or CPU
One small binary with no dependencies, signed build provenance on every release
One local search over your documents and photos: type what you are looking for and get matching files and pictures together, on any laptop or an NVIDIA GPU.
Our PixaFind work showed how much of local AI search is plumbing that every team rebuilds: parsing, chunking, embedding, indexing, GPU batching. Rush Index packages it as one Rust binary, with nothing leaving the machine. On public benchmarks, hybrid search beats either keyword or vector search alone, and an NVIDIA L4 indexes text about 16 times faster than four CPU cores.
Documents and photos in one ranked list; photos only appear when they genuinely match
A small, cleanly licensed reranker you can reproduce with one command: a 0.6B model taught by a 4B one, on permissively licensed data only.
The best small rerankers are trained on data with non-commercial licences. Rush Small asks how much of a 4B teacher a 0.6B student can learn from Apache-2.0 data alone, and publishes the whole recipe and every number. v0 matches the base model on average across five BEIR test sets, gains 3 points on finance questions and gives a little back on argument search. The next version targets a broader data mix.
Distillation from Qwen3-Reranker-4B on NVIDIA GPUs, in one reproducible run
Evaluated on five public BEIR test sets against the base model and the teacher
GGUF export that serves as a reranker in llama.cpp, Ollama and rerankd
Every model and dataset pinned to a commit; licences stated for all data
