# Tokz Labs > Tokz Labs is an AI lab building small, fast models for AI agents. Fine Tune replaces an LLM call that picks from a short list with a small model you own, run on our servers or free on your machine. The Tokz Decisions API (POST /v1/decide) answers typed questions about text with open OpenDecision models, on one API key and one prepaid balance. Every account starts with $5 of free credit at signup: No card, no expiry. 200M billed tokens on Large-Packed. The decision engine returns the best option and a probability for every option, from the same Apache-2.0 OpenDecision weights published on Hugging Face. Tokz processes the text you send to /decide only for the duration of a request, and does not store or log it. ## Fine Tune - [Fine Tune](https://tokz.si/fine-tune): Replace an LLM call with a small model you own. Fine Tune trains your own copy of OpenDecision-Small on your text and one question, then scores both models on the same rows. A Word file of 2,300 texts, spam or not spam, is labeled by AI and trains one small model on a GPU, spam-filter. It scores 99.2 macro-F1, labeled by AI, against 48.3 for OpenDecision-Small on the same 240 held-out rows. The model is 31 MB, answers in about 42 ms on one CPU thread, and marks a spam text from the file as spam at 99.9999%. - [Fine Tune docs](https://docs.tokz.si/fine-tune): Create, run, download and delete a model; prices, limits and errors Fine Tune prices: - Fine tune with a file: $2.00. Only if it beats OpenDecision-Small on rows you labeled. - Fine tune with no file: $0. No charge. - Hosted calls: $0.025. Per 1M billed tokens, at least 256 tokens a request, the same as OpenDecision-Small. - Local calls and downloads: $0. Free. Starting any fine tune needs $2.00 in your balance. ## API - [Tokz Decisions API](https://tokz.si/decisions): Ask typed questions about text and get a probability for every option. POST /v1/decide. Up to sixteen questions on up to sixty-four texts per request. No training run. Question types: choice, tags, yes / no, score. Models: OpenDecision-Small (71M parameters), OpenDecision-Large-Packed (434M parameters), OpenDecision-Large (434M parameters). ## Pricing - [Pricing](https://tokz.si/pricing): Fine Tune and the Decisions API on one prepaid balance; Fine Tune prices are listed above. Decisions API, per 1M billed tokens: OpenDecision-Small $0.025, OpenDecision-Large-Packed $0.025, OpenDecision-Large $0.035. Text and question tokens once per 16 options, plus option names; 3× above 16 options. At least 256 tokens per request. Answers and failed requests are free. $5 of free credit at signup, no card. No seats, no subscription. ## Research - [OpenDecision: Fast Schema-Conditioned Decisions with Packed Candidate Encoding (PDF)](https://github.com/tokz-labs/OpenDecision/releases/latest/download/OpenDecision.pdf): Technical report, Oct 5, 2026. Methods, open baselines, and where the models fall short. - [OpenDecision models](https://huggingface.co/Tokz-labs): Large, Large-Packed and Small, Apache-2.0. - [OpenDecision predictions](https://huggingface.co/datasets/Tokz-labs/OpenDecision-predictions): Model predictions behind the reported numbers. - [OpenDecision code](https://github.com/tokz-labs/OpenDecision): Training, evaluation and inference. - [Research index](https://tokz.si/research): Every number on this site can be rerun from what we publish. Suite macro-F1 (AG News, CLINC150, IMDb, Rotten Tomatoes and XNLI; full test sets, same harness): OpenDecision-Small 79.3, OpenDecision-Large-Packed 87.2, OpenDecision-Large 87.9. Open models only. On label sets they weren't trained on, JevK5, SemIf and Laya score higher. GPU-seconds per 1,000 decisions on one NVIDIA L4: OpenDecision-Small 17.5, OpenDecision-Large-Packed 69.1, OpenDecision-Large 77.4; 4B LLM-based classifier (SemIf) 275. ## Blog - [Introducing Fine Tune: replace an LLM call with a small model you own](https://tokz.si/blog/introducing-fine-tune): 2026-11-09. Upload your file and get back a small model that answers your LLM's question on a CPU. You pay only if it wins. - [Tokz Labs joins the NVIDIA Inception Program](https://tokz.si/blog/tokz-labs-joins-nvidia-inception): 2026-10-30. We are now a member of the NVIDIA Inception program, and here is what we are building next. - [Introducing OpenDecision](https://tokz.si/blog/introducing-opendecision): 2026-10-05. Open encoder models that route, tag and score text, with a probability for every option. Now on the Tokz API. - [RSS feed](https://tokz.si/blog/rss.xml) ## Docs - [Documentation](https://docs.tokz.si): SDK reference and API - [Quickstart](https://docs.tokz.si/quickstart) - [Fine Tune docs](https://docs.tokz.si/fine-tune) - [Decisions API reference](https://docs.tokz.si/decisions) - [TypeScript SDK](https://docs.tokz.si/sdk/typescript): @tokz/sdk on npm - [Python SDK](https://docs.tokz.si/sdk/python): tokz on PyPI - [Billing](https://docs.tokz.si/billing) - [App](https://app.tokz.si): API keys, usage, credit and a playground - [Full text of this site for LLMs](https://tokz.si/llms-full.txt) ## Optional - [Tokz CLI](https://tokz.si/cli): Tokz reads your coding agents' logs and shows where the money went. npx @tokz/cli - [Privacy](https://tokz.si/privacy) - [Terms](https://tokz.si/terms)