# OpenMayhem

> OpenMayhem is a pay-per-use AI inference API over a decentralised provider network, built by Trac Systems and released under the MIT license. The API is OpenAI-compatible: point an existing SDK at https://api.openmayhem.ai/v1 with an API key, pick a model from the live catalog, and pay per request from prepaid credit. Every charge settles against a receipt signed by the provider that did the work.

Key facts:

- Base URL: https://api.openmayhem.ai/v1 — authenticate with `Authorization: Bearer sk-om-v1-...`. Most OpenAI clients work by changing only the base URL and key.
- Endpoints: chat/completions, completions, responses, embeddings (text); tokenize and count_tokens (exact native token counts for supported live model runtimes, never estimated or billed); images/generations, videos, audio/speech, audio/transcriptions, audio/generations, music/generations, workflows (async media jobs with private 12-hour artifact caching and no artifact backups); models (public, cursor-paginated catalog with prices, parameters, provider availability, trust and capacity filters); requests/{id} (settled cost and usage); jobs and artifacts for polling and downloads.
- Billing is reserve-then-settle from prepaid credit: the worst case is held before dispatch, the actual cost charged after, so an account can never spend below zero. Exact USD cost is returned with each response.
- API keys support per-key controls: rate limit, USD spend limits (daily/weekly/monthly/lifetime), a per-request cost cap, expiry, model allowlists and endpoint scopes.
- Paid requests accept provider constraints for minimum attestation tier and context, quantization, strict preferred-provider order, hedging, native-unit throughput and bounded admission wait. Context and hedging are text-only; the remaining applicable controls work across endpoint families. Async jobs retain these controls through retries and reconciliation.
- Catalog query filters: `live`, `endpoint_family`, `modality`, `attestation_tier`, `min_attestation_tier`, `max_input_price_per_million_usd`, `max_output_price_per_million_usd`, `max_representative_request_price_usd`, `min_context_tokens`, `quantization`, `tools`, `structured_output`, `provider`, `prompt_confidential`, `sort`, `order`, `limit`, and opaque `cursor`.
- Public routing headers: `X-Mayhem-Min-Att-Tier`, `X-Mayhem-Min-Ctx`, `X-Mayhem-Quant`, `X-Mayhem-Prefer-Providers`, `X-Mayhem-Hedge`, `X-Mayhem-Min-Tok-S`, `X-Mayhem-Max-Wait-Ms`, `X-OpenMayhem-Max-Cost-Usd`, `X-OpenMayhem-Max-Input-Price-Per-Million-Usd`, `X-OpenMayhem-Max-Output-Price-Per-Million-Usd`, and `X-OpenMayhem-Max-Representative-Request-Price-Usd`.
- `X-OpenMayhem-Max-Cost-Usd` caps the customer's worst-case retail hold. It is separate from the private wholesale `X-Mayhem-Max-Price-Au`, which hosted customers cannot control.
- The three price ceiling headers cap the per-unit retail price accepted for the model of the request: they take the same numbers as the `max_input_price_per_million_usd`, `max_output_price_per_million_usd` and `max_representative_request_price_usd` catalog filters, and a higher admitted price fails with 403 `price_ceiling_exceeded` before any credit is held.
- Errors use a JSON envelope with stable machine-readable codes; 429 carries Retry-After.
- Media settings are capability-dependent: inspect request_contracts, workflow, workflow_media and live_capacity_profiles from the public model catalog. Approved LoRAs, references, presets, custom dimensions and batches are available only where supported; reference uploads do not install LoRA weights. See [Media generation](https://openmayhem.ai/docs/media.md).
- Media example prices illustrate specific settings. Studio estimates describe expected charges; maximum credit holds can be higher and unused credit is released after settlement. Representative-price filters and ceilings use the tariff reference, not the example image price. See [Billing and receipts](https://openmayhem.ai/docs/billing.md).
- Providers earn by running open models on their own hardware; the network routes each request to a live provider and rejects it before any charge when none is available.
- Proxied sellers resell a model they already run. A proxied seller pays a one-time $10 admission fee in TNK, USD or TAP the first time they open or join a proxied market, and every later market for the same seller is free. Native providers never pay. Native and proxied markets are separate, and a request never moves between them.

## API documentation (markdown)

The pages below are served as plain markdown for programmatic consumption.
The complete set in one document: [llms-full.txt](https://openmayhem.ai/llms-full.txt).
Machine-readable OpenAPI 3.1 spec: [openapi.json](https://openmayhem.ai/openapi.json).

- [Introduction](https://openmayhem.ai/docs/introduction.md): What OpenMayhem is and how the API is priced.
- [Quickstart](https://openmayhem.ai/docs/quickstart.md): From account to first billed request in five minutes.
- [Authentication](https://openmayhem.ai/docs/authentication.md): API keys, scopes, spend limits and per-key controls.
- [Endpoint reference](https://openmayhem.ai/docs/reference.md): Every endpoint, its scope and its purpose, on one page.
- [Routing and trust](https://openmayhem.ai/docs/routing.md): Catalog filters, provider constraints, attestation and retail cost limits.
- [Proxy models](https://openmayhem.ai/docs/proxy.md): Provider offers and saved routing profiles in API, Studio and MCP.
- [Chat completions](https://openmayhem.ai/docs/chat.md): OpenAI-compatible text generation and embeddings.
- [Streaming](https://openmayhem.ai/docs/streaming.md): Opt-in progressive SSE and optional Chat/Completions usage.
- [Media generation](https://openmayhem.ai/docs/media.md): Images, video, speech, music and workflows as jobs.
- [Errors](https://openmayhem.ai/docs/errors.md): The error envelope and every status you can receive.
- [OpenMayhem Studio](https://openmayhem.ai/docs/studio.md): Chat, build and create with live models in your browser.
- [MCP for agents](https://openmayhem.ai/docs/mcp.md): Connect MCP clients to live models with explicit cost limits.
- [Billing and receipts](https://openmayhem.ai/docs/billing.md): Reserve-then-settle credit and signed receipts.
- [Sell on OpenMayhem](https://openmayhem.ai/docs/sell.md): Resell a model you run as a proxied seller, or serve open models natively.

## Resources

- [Models and pricing](https://openmayhem.ai/models): live catalog of the hosted API with per-model prices, provider availability, and measured performance
- [API and MCP](https://openmayhem.ai/developers): the OpenAI-compatible API and the MCP server for AI apps, with routing headers, MCP arguments and the endpoint and tool list
- [Earn](https://openmayhem.ai/earn): sell access to a model you run as a proxied seller, or serve open models natively on your own hardware
- [Playground](https://openmayhem.ai/playground): try chat, image, and speech models in the browser
- [Studio guide](https://openmayhem.ai/docs/studio.md): browser projects, Chat, and a step-by-step first media generation
- [Explore models](https://openmayhem.ai/explore): browsable catalog of open models with capabilities, licenses, and hardware requirements
- [API documentation](https://openmayhem.ai/docs): human-readable docs (each page also available as .md)
- [GitHub repository](https://github.com/Trac-Systems/openmayhem): source code, setup guide, and issue tracker
- [Community on Reddit](https://www.reddit.com/r/Open_Mayhem/): questions and discussion
- [Community on Telegram](https://t.me/openmayhem): community chat and updates
- [Trac Systems](https://www.tracsystems.io/): the company behind OpenMayhem

## Available models

- [Qwen3.6 35B A3B Uncensored](https://openmayhem.ai/explore/models/qwen3-6-35b-a3b-uncensored): A fast multimodal Qwen model with fewer refusal behaviors. Runs from 24 GB GPU memory. License: Apache-2.0.
- [Huihui Agents A1](https://openmayhem.ai/explore/models/agents-a1-abliterated-q4-k): A long-horizon agent model with fewer refusal behaviors in a local Q4_K build. Runs from 24 GB GPU memory. License: Apache-2.0.
- [Ternary Bonsai 27B](https://openmayhem.ai/explore/models/ternary-bonsai-27b): A 27B multimodal reasoning model compressed into a laptop-sized ternary build. Runs from CPU · 16 GB RAM. License: Apache-2.0.
- [Gemma 4 E4B](https://openmayhem.ai/explore/models/gemma-4-e4b): A compact multimodal model designed for everyday hardware. Runs from CPU · 16 GB RAM. License: Apache-2.0.
- [GPT-OSS 20B](https://openmayhem.ai/explore/models/gpt-oss-20b): OpenAI's smaller open-weight reasoning model for tools and agents. Runs from CPU · 24 GB RAM. License: Apache-2.0.
- [Qwen3.6 35B A3B](https://openmayhem.ai/explore/models/qwen3-6-35b-a3b): An efficient multimodal model for capable chat and agent workflows. Runs from 24 GB GPU memory. License: Apache-2.0.
- [Ornith 1.0 9B](https://openmayhem.ai/explore/models/ornith-1-0-9b): A compact reasoning model tuned for agentic tasks and tool use. Runs from 8 GB GPU memory. License: MIT.
- [Devstral Small 2 24B](https://openmayhem.ai/explore/models/devstral-small-2-24b): A software-engineering model built for codebases and coding agents. Runs from 16 GB GPU memory. License: Apache-2.0.
- [GPT-OSS 120B](https://openmayhem.ai/explore/models/gpt-oss-120b): OpenAI's larger open-weight model for advanced reasoning and tools. Runs from 80 GB GPU memory. License: Apache-2.0.
- [Qwen3.5 9B Uncensored](https://openmayhem.ai/explore/models/qwen3-5-9b-uncensored): A smaller multimodal Qwen variant with fewer refusal behaviors. Runs from 8 GB GPU memory. License: Apache-2.0.
- [Qwythos 9B MTP](https://openmayhem.ai/explore/models/qwythos-9b-mtp): A long-context 9B model for creative and instruction-heavy work. Runs from 8 GB GPU memory. License: Apache-2.0.
- [Z-Image Turbo](https://openmayhem.ai/explore/models/z-image-turbo): A fast image generator that can run on consumer graphics cards. Runs from 16 GB GPU memory. License: Apache-2.0.
- [FLUX.2 Klein 4B](https://openmayhem.ai/explore/models/flux-2-klein-4b): A compact FLUX model for quick image generation and editing. Runs from 16 GB GPU memory. License: Apache-2.0.
- [Chroma1 HD](https://openmayhem.ai/explore/models/chroma1-hd): An Apache-licensed image model for detailed, open-ended generation. Runs from 24 GB GPU memory. License: Apache-2.0.
- [Qwen Image 2512](https://openmayhem.ai/explore/models/qwen-image-2512): A larger image model with strong text rendering and prompt fidelity. Runs from 24 GB GPU memory. License: Apache-2.0.
- [Sulphur-2](https://openmayhem.ai/explore/models/sulphur-2): An uncensored LTX-2.3 video model that generates motion and synchronized audio. Runs from 48 GB GPU memory. License: Not declared.
- [Chatterbox](https://openmayhem.ai/explore/models/chatterbox): An expressive English speech model with zero-shot voice cloning from Resemble AI. Runs from 6 GB GPU memory. License: MIT.
- [Kokoro 82M](https://openmayhem.ai/explore/models/kokoro-82m): A tiny, fast speech model with a broad set of ready-made voices. Runs from CPU · 4 GB RAM. License: Apache-2.0.
- [Whisper Large V3 Turbo](https://openmayhem.ai/explore/models/whisper-large-v3-turbo): A multilingual speech recognizer optimized for faster transcription. Runs from CPU · 8 GB RAM. License: MIT.
- [Parakeet TDT 0.6B V3](https://openmayhem.ai/explore/models/parakeet-tdt-0-6b-v3): A compact multilingual transcription model built for high throughput. Runs from CPU · 4 GB RAM. License: CC-BY-4.0.
- [ACE-Step 1.5](https://openmayhem.ai/explore/models/ace-step-1-5): An open music model for full songs, vocals, and instrumental tracks. Runs from 4 GB GPU memory. License: MIT.
- [BGE-M3](https://openmayhem.ai/explore/models/bge-m3): A versatile multilingual embedding model for several retrieval styles. Runs from CPU · 8 GB RAM. License: MIT.
- [Qwen3 Embedding 0.6B](https://openmayhem.ai/explore/models/qwen3-embedding-0-6b): A compact multilingual embedding model with flexible dimensions. Runs from CPU · 8 GB RAM. License: Apache-2.0.
- [BGE Reranker V2 M3](https://openmayhem.ai/explore/models/bge-reranker-v2-m3): A multilingual reranker that improves the order of search results. Runs from CPU · 8 GB RAM. License: Apache-2.0.
- [GLM-5.2 Abliterated](https://openmayhem.ai/explore/models/glm-5-2-abliterated): A very large agentic GLM variant for validated multi-GPU clusters. Runs from 4+ GPUs · 253 GB total. License: MIT.

## Model catalog FAQ

### What are open-weight AI models?

Open-weight models publish their trained weights so anyone can download, inspect, and run them. Every model in this catalog links back to its source repository, where you can read the model card, license, and creator guidance.

### Do I need my own GPU to use these models?

No. When you use a model through OpenMayhem, providers on the marketplace run the hardware and you pay per request through an OpenAI-compatible API. The hardware requirements shown in the catalog matter when you want to earn money by serving a model as a provider.

### What does fewer refusals mean?

Models marked with fewer refusals are community-modified variants of open models with reduced refusal behavior, often called abliterated. Each one is labeled in the catalog, credits the modification author, and links to the modified repository so you can review exactly what changed.

### Can I use these models commercially?

Most models in the catalog use permissive licenses like Apache 2.0 and MIT, which allow commercial use. Some use CC BY, which requires attribution. Every model page shows its license and links to the source repository where the full terms live.

### How do models get added to the catalog?

The catalog lists the models that providers can serve on the marketplace. If you want a model that is not listed yet, you can request it in the community and it will be considered for onboarding.