NANITES LABORATORIES

AI NANOBOT FINE-TUNING LABORATORY ยท EST. 2026

Nanites Logo

The Premise

Modern foundation models are capable of an extraordinary range of things, but not every interesting problem calls for a general-purpose system. A nanobot can instead be given a narrower purpose, a particular body of knowledge, a recognizable manner, or a specialized way of working. Fine-tuning gives us a way to make those small changes deliberately.

Nanites is interested in what happens at that scale.

The laboratory develops and releases compact, specialized AI models built from existing foundations and adapted through focused training. Rather than attempting to produce a single model for every possible task, each nanobot is treated as its own experiment. Some are highly specific. Others explore a particular character, domain, medium, or behavior. The common thread is that they are deliberately made for something.

This makes Nanites as much a workshop as a research project. We collect material, prepare training sets, experiment with fine-tuning methods, evaluate the results, and keep the machines that turn out to be interesting. The models are kept small where practical, making them easier to experiment with, modify, and run.

The work is intentionally open-ended. A useful nanobot is worth keeping, but usefulness is not the only measure. A model that captures something unusual, demonstrates an unexpected behavior, or simply makes for an interesting experiment has a place in the laboratory too.

The name comes from the machines themselves: small systems, each adapted for a particular job, capable of being made in many different forms.

Organization

Org Chart

The laboratory operates as a loose collective of nanobots โ€” each a complete, self-contained fine-tuning pipeline with its own data acquisition, preprocessing, training, evaluation, and publishing pipeline. There is no central model; each nanobot is an independent research artifact.

COMPUTE Magazine

Our "publications" are the model cards, datasets, and released weights on Hugging Face. Each nanobot ships with full provenance: training data sources, hyperparameters, compute budget, and evaluation results.

Two Axis Plot

Methodology map: each nanobot lives somewhere on the spectrum of domain specificity vs. modality (text/image/audio), with compute budget as the third axis. We aim for the lower-right quadrant: highly specific, multimodal, low compute.

Released Nanobots

๐ŸŽจ Thangka โ€” Tibetan Thangka Painting (Image)

Thangka

Nanite-Labs/nanites-thangka-sd-1.5 on Hugging Face. Stable Diffusion 1.5 + LoRA (r=16), trained on 85 curated thangka paintings (1,200 steps @ 8GB VRAM). Captures the iconography, palette, and compositional grammar of Tibetan Buddhist thangka paintings. A curated v0.2 set with verified provenance is in progress.

๐Ÿ‹ Whaler โ€” 19th-Century Whaling Logbooks (Text)

Whaler

Nanite-Labs/nanites-whaler-1b-chat on Hugging Face. Llama-3.2-1B-Instruct + QLoRA (r=16, unsloth) on 220 Q&A pairs distilled from 19th-century whaling logbooks (Devstral-24B teacher). A chat model that answers as a period-appropriate first mate โ€” the companion you want on watch.

๐Ÿ“œ Medieval โ€” 15th-Century English Correspondent (Text)

Medieval

Nanite-Labs/nanites-medieval-e2b-chat on Hugging Face. Gemma-4-E2B + QLoRA (r=16) fine-tuned on real pre-1500 English sources: CEECS letters (Oxford), Paston Letters, Chaucer, Gower, Langland, Gawain, Lydgate, Hoccleve. A chat model that speaks as a literate 15th-century English correspondent โ€” estates, lawsuits, wool trade, war in France, household management, legal disputes. Multi-turn conversation supported via Gemma chat template.

๐Ÿชจ Petroglyph โ€” Ancient Rock Art (Image)

Petroglyph

Nanite-Labs/nanites-petroglyph-sdxl-1.0 on Hugging Face. SDXL + LoRA trained on photographed petroglyph sites across the American Southwest โ€” anthropomorphs, spirals, bighorn sheep, celestial markers. Dataset curated from site photography across multiple cultures and millennia.

๐ŸŽต Smoky โ€” Smoky Mountain English Dialect (Speech + Text Collection)

Smoky

Nanite-Labs/nanites-smoky-speech-earl and Nanite-Labs/nanites-smoky-speech-ethel on Hugging Face (CosyVoice2-0.5B gender-pooled SFT โ€” first generation). Nanite-Labs/nanites-smoky-speech-otis and Nanite-Labs/nanites-smoky-speech-opal on Hugging Face (Fun-CosyVoice3-0.5B gender-pooled SFT on fresh random speaker pools โ€” second generation). Nanite-Labs/nanites-smoky-1b-chat on Hugging Face (Llama-3.2-1B + QLoRA on Joseph Hall 1939 transcripts).

Smoky 8-track

Smoky Cassette

A regional dialect collection, not a voice clone. Gender-pooled CosyVoice voices โ€” first generation Earl (male) and Ethel (female), second generation Otis (male) and Opal (female, Fun-CosyVoice3) โ€” are anonymized pooled embeddings; no individual speaker is exposed. The chat model provides text generation in the regional register. The Smoky collection: speech models, chat model, and vintage media artifacts (8-track, cassette, VHS) together form the Smoky collection.

๐Ÿ–ฅ๏ธ Retro Web Designer โ€” Pre-Y2K Web Pages + Assets (Code + Image)

Retro Web Designer

Nanite-Labs/nanites-retro-web-e4b-designer on Hugging Face โ€” one repo, two LoRAs. A gemma-4-E4B-it QLoRA (r=16) fine-tuned on harvested pre-Y2K web source code writes period-correct 1994โ€“1999 web pages โ€” tables, spacer GIFs, marquee and blink, hit counters, guestbooks, frames, image maps, tiled backgrounds, web-ring links, "under construction" banners, era typography. No modern CSS (flexbox/grid), no SPA frameworks, no modern JS. Its Stable Diffusion 1.5 LoRA (r=16) on ~1,500 curated GIFCities-era GIFs materializes every dangling asset reference โ€” buttons, banners, dividers, backgrounds, clipart โ€” so a generated page renders fully local with on-era images. Trigger tokens retro web 1999 (code) / retroweb90s (assets); loaded together in one process on 8GB.

Example pages generated end-to-end (code + synthesized assets) ship in the samples/ folder of the repo.

๐Ÿฆ Fiorino โ€” 1400s Ledgers to Tabular Foundation (Tabular)

Fiorino

Nanite-Labs/nanites-fiorino-tabular on Hugging Face โ€” INTRODUCING FIORINO, nano quattrocento tabular fondamento model with synthetic and real-data priors. A 4.5M-parameter in-context learner: labeled context rows in, query predictions out, one forward pass, no fine-tuning. Pretrained on pre-1600 merchant-bank ledgers (Borromei, Crown loans, Bristol customs, Irish Exchequer) plus a world-profiled synthetic diet. TabArena-38 cls 0.818 (27/38 vs RF); reg +0.619 with first wins banked. 226,000 tables vs 100,000,000+ at the big labs. Sister code: github.com/frankiethull/fiorino (python/ + r/).

Methodology

Nanites Org Card

Hardware reality: Single RTX 4070 (8GB VRAM). Every pipeline must fit. We use:

Data discipline: Public domain or locally licensed only. No redistributed copyrighted transcripts. Teacher distillation (Devstral-24B, Nemotron-3-nano) for Q&A expansion. Manual curation > scale.

Evaluation: Human-in-the-loop for audio/image. Automated WER for ASR. Held-out validation for all. No self-approval via VLMs.

This page is the public workbench for the laboratory. Models and experiments will be added here as they are completed, along with the datasets and other materials that make them possible.

For now, the laboratory is small, the machines are compact, and the experiments are underway.

Late-Nite Sales

Nanites Laboratories
AI nanobots, fine-tuned for particular purposes.