# Onrup

> Fine-tune open-weight models, prove they beat the baseline, and serve them on an OpenAI-compatible endpoint. Serverless compute across multiple providers.

Onrup is an API-first platform for fine-tuning open-weight language models and serving them on an OpenAI-compatible endpoint. Compute is serverless across multiple providers; a caller chooses a GPU class and a scaling mode rather than a machine. Three properties define the product: cost is authorised and reserved before compute is leased, a blinded evaluation gate must pass before a model can be deployed or published, and trained weights are always downloadable.

Deliberately not published: which compute providers are used, how capacity is selected between them, and how the platform is deployed. Everything a caller can observe — GPU class, memory, hourly rate, scaling behaviour, what is metered and when — is published in full and linked below.

Contact: hello@onrup.com

## Start here

- [How it works](https://www.onrup.com/how-it-works): dataset to deployed model in five API calls.
- [Pricing](https://www.onrup.com/pricing): published rate card, 13 GPU classes, from $0.09 per GPU-hour.
- [Quickstart](https://www.onrup.com/docs/quickstart): every call, in full.
- [FAQ](https://www.onrup.com/faq): including the questions where the answer is no.

## What makes it different

- [Predictable cost](https://www.onrup.com/predictable-cost): spend is reserved against a limit before compute is leased, not alerted on afterwards.
- [Your weights, your endpoint](https://www.onrup.com/your-weights): download or publish any finished run; serving speaks the OpenAI API.
- [The evaluation gate](https://www.onrup.com/evaluation-gate): blinded comparison against the incumbent; pass, fail or inconclusive.
- [Serverless compute](https://www.onrup.com/serverless-compute): a GPU class, not a machine, across multiple providers.

## Documentation

- [Quickstart](https://www.onrup.com/docs/quickstart): How do I fine-tune and deploy a model on Onrup?
- [Authentication](https://www.onrup.com/docs/authentication): How does authentication work on the Onrup API?
- [Datasets](https://www.onrup.com/docs/datasets): What dataset formats does Onrup accept and how are they validated?
- [Runs](https://www.onrup.com/docs/runs): How does a training run work on Onrup?
- [Evaluation and the gate](https://www.onrup.com/docs/evaluation): How does the Onrup evaluation gate decide whether a model can be deployed?
- [Deployments](https://www.onrup.com/docs/deployments): How do I serve a fine-tuned model on Onrup?
- [Publishing models](https://www.onrup.com/docs/publishing): How do I export or publish a model trained on Onrup?
- [Long-running operations](https://www.onrup.com/docs/operations): How do I track a long-running job on the Onrup API?
- [Billing and spend limits](https://www.onrup.com/docs/billing): How does Onrup billing work?
- [Errors](https://www.onrup.com/docs/errors): What error codes does the Onrup API return?
- [Rate limits and quotas](https://www.onrup.com/docs/limits): What are the Onrup rate limits and quotas?
- [Security model](https://www.onrup.com/docs/security): How does Onrup isolate and protect tenant data?
- [API reference](https://www.onrup.com/docs/api): What endpoints does the Onrup API expose?

## Model catalogue (43 models, 12 families)

- [SmolLM2 1.7B](https://www.onrup.com/models/smollm2-1.7b): 1.7B, 6 GB LoRA / 4 GB QLoRA, Apache 2.0.
- [SmolLM3 3B](https://www.onrup.com/models/smollm3-3b): 3B, 10 GB LoRA / 6 GB QLoRA, Apache 2.0.
- [Qwen3 0.6B](https://www.onrup.com/models/qwen3-0.6b): 0.6B, 4 GB LoRA / 3 GB QLoRA, Apache 2.0.
- [Qwen3 1.7B](https://www.onrup.com/models/qwen3-1.7b): 1.7B, 6 GB LoRA / 4 GB QLoRA, Apache 2.0.
- [Qwen3 4B](https://www.onrup.com/models/qwen3-4b): 4B, 12 GB LoRA / 8 GB QLoRA, Apache 2.0.
- [Qwen3 8B](https://www.onrup.com/models/qwen3-8b): 8B, 18 GB LoRA / 14 GB QLoRA, Apache 2.0.
- [Qwen3 14B](https://www.onrup.com/models/qwen3-14b): 14B, 28 GB LoRA / 20 GB QLoRA, Apache 2.0.
- [Qwen3 32B](https://www.onrup.com/models/qwen3-32b): 32B, 64 GB LoRA / 48 GB QLoRA, Apache 2.0.
- [Qwen3 30B-A3B](https://www.onrup.com/models/qwen3-30b-a3b): 30B, 64 GB LoRA / 36 GB QLoRA, Apache 2.0.
- [Llama 3.2 1B](https://www.onrup.com/models/llama-3.2-1b): 1B, 4 GB LoRA / 3 GB QLoRA, Llama 3.2 Community.
- [Llama 3.2 3B](https://www.onrup.com/models/llama-3.2-3b): 3B, 10 GB LoRA / 6 GB QLoRA, Llama 3.2 Community.
- [Llama 3.1 8B](https://www.onrup.com/models/llama-3.1-8b): 8B, 18 GB LoRA / 14 GB QLoRA, Llama 3.1 Community.
- [Llama 3.3 70B](https://www.onrup.com/models/llama-3.3-70b): 70B, 140 GB LoRA / 48 GB QLoRA, Llama 3.3 Community.
- [Llama 4 Scout 109B](https://www.onrup.com/models/llama-4-scout-109b): 109B, 220 GB LoRA / 80 GB QLoRA, Llama 4 Community.
- [Mistral 7B v0.3](https://www.onrup.com/models/mistral-7b-v0.3): 7B, 16 GB LoRA / 12 GB QLoRA, Apache 2.0.
- [Mistral Nemo 12B](https://www.onrup.com/models/mistral-nemo-12b): 12B, 24 GB LoRA / 16 GB QLoRA, Apache 2.0.
- [Mistral Small 3 24B](https://www.onrup.com/models/mistral-small-3-24b): 24B, 48 GB LoRA / 36 GB QLoRA, Apache 2.0.
- [Phi-4 Mini 3.8B](https://www.onrup.com/models/phi-4-mini): 3.8B, 12 GB LoRA / 8 GB QLoRA, MIT.
- [Phi-4 14B](https://www.onrup.com/models/phi-4-14b): 14B, 28 GB LoRA / 20 GB QLoRA, MIT.
- [Gemma 3 1B](https://www.onrup.com/models/gemma-3-1b): 1B, 4 GB LoRA / 3 GB QLoRA, Gemma Terms of Use.
- [Gemma 3 4B](https://www.onrup.com/models/gemma-3-4b): 4B, 12 GB LoRA / 8 GB QLoRA, Gemma Terms of Use.
- [Gemma 3 12B](https://www.onrup.com/models/gemma-3-12b): 12B, 24 GB LoRA / 18 GB QLoRA, Gemma Terms of Use.
- [Gemma 3 27B](https://www.onrup.com/models/gemma-3-27b): 27B, 56 GB LoRA / 36 GB QLoRA, Gemma Terms of Use.
- [DeepSeek-R1 Distill Qwen 7B](https://www.onrup.com/models/deepseek-r1-distill-qwen-7b): 7B, 16 GB LoRA / 12 GB QLoRA, Apache 2.0.
- [DeepSeek-R1 Distill Llama 8B](https://www.onrup.com/models/deepseek-r1-distill-llama-8b): 8B, 18 GB LoRA / 14 GB QLoRA, Llama 3.1 Community.
- [LFM2 350M](https://www.onrup.com/models/lfm2-350m): 0.35B, 3 GB LoRA / 2 GB QLoRA, LFM Open.
- [LFM2 700M](https://www.onrup.com/models/lfm2-700m): 0.7B, 4 GB LoRA / 3 GB QLoRA, LFM Open.
- [LFM2 1.2B](https://www.onrup.com/models/lfm2-1.2b): 1.2B, 5 GB LoRA / 4 GB QLoRA, LFM Open.
- [LFM2 2.6B](https://www.onrup.com/models/lfm2-2.6b): 2.6B, 8 GB LoRA / 6 GB QLoRA, LFM Open.
- [LFM2 8B-A1B](https://www.onrup.com/models/lfm2-8b-moe): 8.3B, 18 GB LoRA / 12 GB QLoRA, LFM Open.
- [LFM2 24B-A2B](https://www.onrup.com/models/lfm2-24b-a2b): 24B, 48 GB LoRA / 28 GB QLoRA, LFM Open.
- [Granite 3.0 2B](https://www.onrup.com/models/granite-3.0-2b): 2B, 7 GB LoRA / 5 GB QLoRA, Apache 2.0.
- [Granite 3.0 8B](https://www.onrup.com/models/granite-3.0-8b): 8B, 18 GB LoRA / 14 GB QLoRA, Apache 2.0.
- [Granite 4.1 8B](https://www.onrup.com/models/granite-4.1-8b): 8B, 18 GB LoRA / 14 GB QLoRA, Apache 2.0.
- [gpt-oss 20B](https://www.onrup.com/models/gpt-oss-20b): 20.9B, 40 GB LoRA / 24 GB QLoRA, Apache 2.0.
- [gpt-oss 120B](https://www.onrup.com/models/gpt-oss-120b): 116.8B, 220 GB LoRA / 80 GB QLoRA, Apache 2.0.
- [Falcon 3 1B](https://www.onrup.com/models/falcon-3-1b): 1B, 4 GB LoRA / 3 GB QLoRA, TII Falcon LLM.
- [Falcon 3 3B](https://www.onrup.com/models/falcon-3-3b): 3B, 10 GB LoRA / 6 GB QLoRA, TII Falcon LLM.
- [Falcon 3 7B](https://www.onrup.com/models/falcon-3-7b): 7B, 16 GB LoRA / 12 GB QLoRA, TII Falcon LLM.
- [Falcon 3 10B](https://www.onrup.com/models/falcon-3-10b): 10B, 22 GB LoRA / 16 GB QLoRA, TII Falcon LLM.
- [OLMo 2 7B](https://www.onrup.com/models/olmo-2-7b): 7B, 16 GB LoRA / 12 GB QLoRA, Apache 2.0.
- [OLMo 3 7B](https://www.onrup.com/models/olmo-3-7b): 7B, 16 GB LoRA / 12 GB QLoRA, Apache 2.0.
- [OLMo 3 32B](https://www.onrup.com/models/olmo-3-32b): 32B, 64 GB LoRA / 48 GB QLoRA, Apache 2.0.

## GPU classes (13)

- [RTX 3080](https://www.onrup.com/gpus/rtx-3080-12g): 12 GB, $0.09/hr training, $0.11/hr serving.
- [RTX 4000 Ada](https://www.onrup.com/gpus/rtx-4000-ada-20g): 20 GB, $0.09/hr training, $0.11/hr serving.
- [L4](https://www.onrup.com/gpus/l4-24g): 24 GB, $0.17/hr training, $0.20/hr serving.
- [RTX 3090](https://www.onrup.com/gpus/rtx-3090-24g): 24 GB, $0.22/hr training, $0.26/hr serving.
- [RTX 4090](https://www.onrup.com/gpus/rtx-4090-24g): 24 GB, $0.38/hr training, $0.43/hr serving.
- [A40](https://www.onrup.com/gpus/a40-48g): 48 GB, $0.42/hr training, $0.48/hr serving.
- [RTX 6000 Ada](https://www.onrup.com/gpus/rtx-6000-ada-48g): 48 GB, $0.61/hr training, $0.71/hr serving.
- [L40S](https://www.onrup.com/gpus/l40s-48g): 48 GB, $0.78/hr training, $0.90/hr serving.
- [A6000](https://www.onrup.com/gpus/a6000-48g): 48 GB, $0.65/hr training, $0.75/hr serving.
- [A100 40 GB](https://www.onrup.com/gpus/a100-40g): 40 GB, $1.17/hr training, $1.35/hr serving.
- [A100 80 GB](https://www.onrup.com/gpus/a100-80g): 80 GB, $1.68/hr training, $1.94/hr serving.
- [H100 80 GB](https://www.onrup.com/gpus/h100-80g): 80 GB, $2.59/hr training, $2.99/hr serving.
- [H200](https://www.onrup.com/gpus/h200-141g): 141 GB, $4.55/hr training, $5.25/hr serving.

## Guides

- [Should you fine-tune, or is it a prompt problem?](https://www.onrup.com/guides/should-you-fine-tune): When is fine-tuning worth it compared to prompt engineering or retrieval?
- [LoRA, QLoRA or full fine-tuning: how to choose](https://www.onrup.com/guides/lora-vs-qlora-vs-full-fine-tuning): What is the difference between LoRA, QLoRA and full fine-tuning?
- [How much VRAM does fine-tuning actually need?](https://www.onrup.com/guides/how-much-vram-do-i-need): How much GPU memory do I need to fine-tune a model?
- [Preparing a fine-tuning dataset](https://www.onrup.com/guides/preparing-a-fine-tuning-dataset): How do I prepare a dataset for fine-tuning?
- [ShareGPT, ChatML and Alpaca: which dataset format to use](https://www.onrup.com/guides/sharegpt-vs-chatml-vs-alpaca): What is the difference between ShareGPT, ChatML and Alpaca dataset formats?
- [How to choose a base model](https://www.onrup.com/guides/choosing-a-base-model): How do I choose which base model to fine-tune?
- [Designing an evaluation set that actually decides things](https://www.onrup.com/guides/designing-an-evaluation-set): How do I build an evaluation set for a fine-tuned model?
- [How to estimate what a fine-tuning run will cost](https://www.onrup.com/guides/estimating-fine-tuning-cost): How much does it cost to fine-tune an open-weight model?
- [Scale to zero or always warm: choosing a serving mode](https://www.onrup.com/guides/scale-to-zero-or-always-warm): Should my model endpoint scale to zero or stay always warm?
- [Migrating from a closed model API to an open-weight fine-tune](https://www.onrup.com/guides/migrating-from-a-closed-model-api): How do I replace a closed model API with a fine-tuned open-weight model?
- [Why fine-tuned models fail, and how to tell which failure you have](https://www.onrup.com/guides/why-fine-tuned-models-fail): Why did my fine-tuned model get worse instead of better?
- [When to use preference tuning instead of supervised fine-tuning](https://www.onrup.com/guides/when-to-use-preference-tuning): When should I use DPO instead of supervised fine-tuning?
- [Using reinforcement learning to improve reasoning](https://www.onrup.com/guides/reinforcement-learning-for-reasoning): How does GRPO work and when should I use reinforcement learning to fine-tune?
- [Serving many fine-tuned models without paying for each one](https://www.onrup.com/guides/serving-many-fine-tuned-models): How can I serve dozens of fine-tuned models economically?
- [Controlling spend on a fine-tuning platform](https://www.onrup.com/guides/controlling-fine-tuning-spend): How do I stop fine-tuning and inference costs running away?
- [Keeping a fine-tuned model portable](https://www.onrup.com/guides/keeping-a-fine-tuned-model-portable): How do I make sure I can move a fine-tuned model to another provider?

## Use cases

- [Customer support triage](https://www.onrup.com/use-cases/customer-support-triage): How do I fine-tune a model to route and tag support tickets?
- [Structured data extraction](https://www.onrup.com/use-cases/structured-data-extraction): How do I fine-tune a model to return reliable JSON from messy documents?
- [Replacing a frontier API on one task](https://www.onrup.com/use-cases/replacing-a-frontier-api): How do I cut my LLM bill by replacing a frontier model with a fine-tuned open one?
- [Domain question answering](https://www.onrup.com/use-cases/domain-question-answering): Should I fine-tune or use retrieval for questions about my own documents?
- [Code review assistant](https://www.onrup.com/use-cases/code-review-assistant): How do I fine-tune a model on my team’s code review standards?
- [SQL generation over your own schema](https://www.onrup.com/use-cases/sql-generation): How do I fine-tune a model to write correct SQL for my database?
- [Tool-calling agents](https://www.onrup.com/use-cases/tool-calling-agent): How do I fine-tune a model to call my tools reliably?
- [Content moderation](https://www.onrup.com/use-cases/content-moderation): How do I fine-tune a moderation model on my own policy?
- [Summarisation in a house style](https://www.onrup.com/use-cases/summarisation-with-house-style): How do I fine-tune a model to summarise in our specific format?
- [PII redaction](https://www.onrup.com/use-cases/pii-redaction): How do I fine-tune a model to redact personal data in our documents?
- [Reasoning with reinforcement learning](https://www.onrup.com/use-cases/reasoning-with-rl): When should I use GRPO instead of supervised fine-tuning?
- [Translation and localisation](https://www.onrup.com/use-cases/translation-and-localisation): How do I fine-tune a model on our terminology and tone across languages?
- [Classification at volume](https://www.onrup.com/use-cases/classification-at-volume): What is the cheapest way to classify millions of documents with an LLM?
- [On-brand copy generation](https://www.onrup.com/use-cases/on-brand-copy-generation): How do I fine-tune a model to write in our brand voice?
- [Semantic routing between models](https://www.onrup.com/use-cases/semantic-routing): How do I route requests between a small and a large model automatically?
- [Document parsing pipelines](https://www.onrup.com/use-cases/document-parsing-pipelines): How do I fine-tune a model to parse messy PDFs and scans consistently?

## Comparisons

- [The landscape](https://www.onrup.com/landscape): sourced market map; every figure links to the vendor's own page.
- [Onrup vs Together AI](https://www.onrup.com/compare/onrup-vs-together-ai): A broad model API with fine-tuning attached, covering the widest catalogue of open-weight models of anyone in this set.
- [Onrup vs Fireworks AI](https://www.onrup.com/compare/onrup-vs-fireworks-ai): Inference, fine-tuning and evaluation behind one API, with fine-tuned adapters served at the same per-token rate as the base model.
- [Onrup vs Modal](https://www.onrup.com/compare/onrup-vs-modal): Serverless compute for arbitrary Python, billed by the second. Not a fine-tuning product — a substrate you build one on.
- [Onrup vs Baseten](https://www.onrup.com/compare/onrup-vs-baseten): Model deployment and serving with strong operational tooling, compliance posture and cold-start engineering.
- [Onrup vs Replicate](https://www.onrup.com/compare/onrup-vs-replicate): A catalogue of community-published models behind one API, billed per second of compute, with fine-tuning on a subset.
- [Onrup vs OpenAI fine-tuning](https://www.onrup.com/compare/onrup-vs-openai-fine-tuning): Fine-tuning of OpenAI’s own closed models, served only from OpenAI. The default first stop, and the thing most teams are trying to leave.
- [Onrup vs Hugging Face](https://www.onrup.com/compare/onrup-vs-hugging-face): The model hub itself, plus AutoTrain for training and Inference Endpoints for serving. The centre of gravity of the open-weight world.
- [Onrup vs Predibase](https://www.onrup.com/compare/onrup-vs-predibase): A fine-tuning platform built around serving many LoRA adapters from a single GPU. Acquired — predibase.com now redirects to Rubrik.
- [Onrup vs OpenPipe](https://www.onrup.com/compare/onrup-vs-openpipe): Capture production traffic from a large model, then train a small one to replace it on that exact distribution.
- [Onrup vs AWS SageMaker](https://www.onrup.com/compare/onrup-vs-aws-sagemaker): The full machine-learning platform on AWS. Enormously capable, correspondingly heavy, and priced per instance-hour.
- [Onrup vs Google Vertex AI](https://www.onrup.com/compare/onrup-vs-google-vertex-ai): Google Cloud’s machine-learning platform, with tuning for Gemini and for a set of open-weight models.
- [Onrup vs AWS Bedrock](https://www.onrup.com/compare/onrup-vs-aws-bedrock): A managed API over several model vendors, with custom-model fine-tuning and provisioned throughput inside AWS.
- [Onrup vs Anyscale](https://www.onrup.com/compare/onrup-vs-anyscale): Managed Ray. Distributed training and serving for teams that have outgrown a single machine.
- [Onrup vs Lamini](https://www.onrup.com/compare/onrup-vs-lamini): Enterprise fine-tuning with a focus on factual accuracy and hallucination reduction, deployable on-premises.

## Integrations

- [OpenAI Python SDK](https://www.onrup.com/integrations/openai-python): Python.
- [OpenAI Node SDK](https://www.onrup.com/integrations/openai-node): TypeScript.
- [curl and plain HTTP](https://www.onrup.com/integrations/curl): Shell.
- [LangChain](https://www.onrup.com/integrations/langchain): Python.
- [LlamaIndex](https://www.onrup.com/integrations/llamaindex): Python.
- [Vercel AI SDK](https://www.onrup.com/integrations/vercel-ai-sdk): TypeScript.
- [LiteLLM](https://www.onrup.com/integrations/litellm): YAML.
- [Pydantic AI](https://www.onrup.com/integrations/pydantic-ai): Python.
- [n8n](https://www.onrup.com/integrations/n8n): Configuration.
- [Open WebUI](https://www.onrup.com/integrations/open-webui): Configuration.
- [Hugging Face Hub](https://www.onrup.com/integrations/hugging-face-hub): HTTP.
- [Continue](https://www.onrup.com/integrations/continue): JSON.

## Glossary (124 terms)

- [Fine-tuning](https://www.onrup.com/glossary/fine-tuning)
- [LoRA](https://www.onrup.com/glossary/lora)
- [QLoRA](https://www.onrup.com/glossary/qlora)
- [Supervised fine-tuning](https://www.onrup.com/glossary/supervised-fine-tuning)
- [Preference tuning](https://www.onrup.com/glossary/preference-tuning)
- [Reinforcement learning from verifiable rewards](https://www.onrup.com/glossary/reinforcement-learning-from-verifiable-rewards)
- [Full fine-tuning](https://www.onrup.com/glossary/full-fine-tuning)
- [Catastrophic forgetting](https://www.onrup.com/glossary/catastrophic-forgetting)
- [Overfitting](https://www.onrup.com/glossary/overfitting)
- [Epoch](https://www.onrup.com/glossary/epoch)
- [Learning rate](https://www.onrup.com/glossary/learning-rate)
- [Learning rate schedule](https://www.onrup.com/glossary/learning-rate-schedule)
- [Warmup](https://www.onrup.com/glossary/warmup)
- [Batch size](https://www.onrup.com/glossary/batch-size)
- [Gradient accumulation](https://www.onrup.com/glossary/gradient-accumulation)
- [Gradient checkpointing](https://www.onrup.com/glossary/gradient-checkpointing)
- [Rank](https://www.onrup.com/glossary/rank)
- [Target modules](https://www.onrup.com/glossary/target-modules)
- [Adapter](https://www.onrup.com/glossary/adapter)
- [Adapter merging](https://www.onrup.com/glossary/adapter-merging)
- [Training step](https://www.onrup.com/glossary/training-step)
- [Checkpoint](https://www.onrup.com/glossary/checkpoint)
- [Early stopping](https://www.onrup.com/glossary/early-stopping)
- [Optimiser state](https://www.onrup.com/glossary/optimiser-state)
- [Gradient](https://www.onrup.com/glossary/gradient)
- [Loss](https://www.onrup.com/glossary/loss)
- [Perplexity](https://www.onrup.com/glossary/perplexity)
- [Training example](https://www.onrup.com/glossary/training-example)
- [ChatML](https://www.onrup.com/glossary/chatml)
- [ShareGPT format](https://www.onrup.com/glossary/sharegpt-format)
- [Alpaca format](https://www.onrup.com/glossary/alpaca-format)
- [Format adapter](https://www.onrup.com/glossary/format-adapter)
- [Dataset validation](https://www.onrup.com/glossary/dataset-validation)
- [Deduplication](https://www.onrup.com/glossary/deduplication)
- [Held-out set](https://www.onrup.com/glossary/held-out-set)
- [Validation split](https://www.onrup.com/glossary/validation-split)
- [Data contamination](https://www.onrup.com/glossary/data-contamination)
- [Synthetic data](https://www.onrup.com/glossary/synthetic-data)
- [Data provenance](https://www.onrup.com/glossary/data-provenance)
- [Sequence length](https://www.onrup.com/glossary/sequence-length)
- [Truncation](https://www.onrup.com/glossary/truncation)
- [Tokenisation](https://www.onrup.com/glossary/tokenisation)
- [Token](https://www.onrup.com/glossary/token)
- [Special tokens](https://www.onrup.com/glossary/special-tokens)
- [Chat template](https://www.onrup.com/glossary/chat-template)
- [Preference pair](https://www.onrup.com/glossary/preference-pair)
- [Evaluation gate](https://www.onrup.com/glossary/evaluation-gate)
- [Blinded evaluation](https://www.onrup.com/glossary/blinded-evaluation)
- [Baseline model](https://www.onrup.com/glossary/baseline-model)
- [Win rate](https://www.onrup.com/glossary/win-rate)
- [Confidence interval](https://www.onrup.com/glossary/confidence-interval)
- [Statistical significance](https://www.onrup.com/glossary/statistical-significance)
- [LLM-as-judge](https://www.onrup.com/glossary/llm-as-judge)
- [Rubric](https://www.onrup.com/glossary/rubric)
- [Critical regression](https://www.onrup.com/glossary/critical-regression)
- [Regression](https://www.onrup.com/glossary/regression)
- [Benchmark](https://www.onrup.com/glossary/benchmark)
- [Safety evaluation](https://www.onrup.com/glossary/safety-evaluation)
- [Inference](https://www.onrup.com/glossary/inference)
- [OpenAI-compatible API](https://www.onrup.com/glossary/openai-compatible-api)
- [Endpoint](https://www.onrup.com/glossary/endpoint)
- [Scale to zero](https://www.onrup.com/glossary/scale-to-zero)
- [Always warm](https://www.onrup.com/glossary/always-warm)
- [Cold start](https://www.onrup.com/glossary/cold-start)
- [Cooldown](https://www.onrup.com/glossary/cooldown)
- [Replica](https://www.onrup.com/glossary/replica)
- [Latency](https://www.onrup.com/glossary/latency)
- [Time to first token](https://www.onrup.com/glossary/time-to-first-token)
- [Throughput](https://www.onrup.com/glossary/throughput)
- [Batching](https://www.onrup.com/glossary/batching)
- [KV cache](https://www.onrup.com/glossary/kv-cache)
- [Context window](https://www.onrup.com/glossary/context-window)
- [Streaming](https://www.onrup.com/glossary/streaming)
- [Server-sent events](https://www.onrup.com/glossary/server-sent-events)
- [Multi-adapter serving](https://www.onrup.com/glossary/multi-adapter-serving)
- [Quantisation](https://www.onrup.com/glossary/quantisation)
- [Tool calling](https://www.onrup.com/glossary/tool-calling)
- [Structured output](https://www.onrup.com/glossary/structured-output)
- [Schema validity](https://www.onrup.com/glossary/schema-validity)
- [Base model](https://www.onrup.com/glossary/base-model)
- [Instruction-tuned model](https://www.onrup.com/glossary/instruction-tuned-model)
- [Open-weight model](https://www.onrup.com/glossary/open-weight-model)
- [Model licence](https://www.onrup.com/glossary/model-licence)
- [Mixture of experts](https://www.onrup.com/glossary/mixture-of-experts)
- [Active parameters](https://www.onrup.com/glossary/active-parameters)
- [Dense model](https://www.onrup.com/glossary/dense-model)
- [Parameter count](https://www.onrup.com/glossary/parameter-count)
- [Attention](https://www.onrup.com/glossary/attention)
- [Transformer](https://www.onrup.com/glossary/transformer)
- [Feed-forward network](https://www.onrup.com/glossary/feed-forward-network)
- [Reasoning model](https://www.onrup.com/glossary/reasoning-model)
- [Distillation](https://www.onrup.com/glossary/distillation)
- [Retrieval-augmented generation](https://www.onrup.com/glossary/retrieval-augmented-generation)
- [Hallucination](https://www.onrup.com/glossary/hallucination)
- [Refusal](https://www.onrup.com/glossary/refusal)
- [Prompt injection](https://www.onrup.com/glossary/prompt-injection)
- [Temperature](https://www.onrup.com/glossary/temperature)
- [Agent](https://www.onrup.com/glossary/agent)
- [GPU class](https://www.onrup.com/glossary/gpu-class)
- [VRAM](https://www.onrup.com/glossary/vram)
- [GPU-hour](https://www.onrup.com/glossary/gpu-hour)
- [Serverless compute](https://www.onrup.com/glossary/serverless-compute)
- [Cost authorisation](https://www.onrup.com/glossary/cost-authorisation)
- [Spend limit](https://www.onrup.com/glossary/spend-limit)
- [Reservation](https://www.onrup.com/glossary/reservation)
- [Metered billing](https://www.onrup.com/glossary/metered-billing)
- [Usage event](https://www.onrup.com/glossary/usage-event)
- [Entitlement](https://www.onrup.com/glossary/entitlement)
- [Quota](https://www.onrup.com/glossary/quota)
- [Rate limit](https://www.onrup.com/glossary/rate-limit)
- [Durable operation](https://www.onrup.com/glossary/durable-operation)
- [Idempotency](https://www.onrup.com/glossary/idempotency)
- [Resumption](https://www.onrup.com/glossary/resumption)
- [Model weights](https://www.onrup.com/glossary/model-weights)
- [Model publishing](https://www.onrup.com/glossary/model-publishing)
- [Model card](https://www.onrup.com/glossary/model-card)
- [Audit log](https://www.onrup.com/glossary/audit-log)
- [API key](https://www.onrup.com/glossary/api-key)
- [Scope](https://www.onrup.com/glossary/scope)
- [Prompt engineering](https://www.onrup.com/glossary/prompt-engineering)
- [Reward function](https://www.onrup.com/glossary/reward-function)
- [Reward hacking](https://www.onrup.com/glossary/reward-hacking)
- [Reward model](https://www.onrup.com/glossary/reward-model)
- [Reference model](https://www.onrup.com/glossary/reference-model)

## Machine-readable

- [OpenAPI document](https://www.onrup.com/openapi.json)
- [Models as JSON](https://www.onrup.com/api/models.json)
- [GPU classes and rates as JSON](https://www.onrup.com/api/pricing.json)
- [Glossary as JSON](https://www.onrup.com/api/glossary.json)
- [Landscape dataset as JSON](https://www.onrup.com/api/landscape.json)
- [Guides as RSS](https://www.onrup.com/rss.xml)
- Every guide is also served as Markdown at its URL plus `.md`.

## Legal

- [Privacy](https://www.onrup.com/legal/privacy)
- [Terms](https://www.onrup.com/legal/terms)
- [Acceptable use](https://www.onrup.com/legal/acceptable-use)