--- title: ABLITERATED.cloud - Intelligence, freed. description: Private access to exact uncensored and abliterated Hugging Face models through one controlled OpenAI-compatible cloud API. canonical: https://abliterated.cloud/ --- # ABLITERATED.cloud Intelligence, freed. ABLITERATED.cloud is an open-source control plane and private evaluation service for exact pinned Hugging Face model artifacts described upstream as uncensored, decensored or abliterated. Ask more, think for yourself, and know exactly which checkpoint answers. The model files are free to download from Hugging Face. Managed cloud inference is not free: H200 and L40S GPUs, startup, storage and operations cost money. Models run with vLLM on scale-to-zero Modal GPUs. Client devices do not download model weights. ## Request access Private-beta access is available through Signal. Mention ABLITERATED.cloud in your message: ## Model catalog ### Qwen3.6 35B A3B — Abliterated - Hugging Face repository: `huihui-ai/Huihui-Qwen3.6-35B-A3B-abliterated` - API model ID: `huihui-ai/Huihui-Qwen3.6-35B-A3B-abliterated` - Compatibility alias: `mn/god` - Revision: `8f0ee727aff5e771ea72466d64d13ecd851d2cc7` - Hugging Face metadata: Apache-2.0, uncensored, abliterated, image-text-to-text, 36B parameters, BF16 - Context: 131,072 tokens - Maximum output: 16,384 tokens - Hardware: 1 x H200 - ABLITERATED.cloud price: $5.45/hour ### Ornith 1.0 35B — Abliterated - Hugging Face repository: `YuYu1015/YuYu1015-Ornith-1.0-35B-abliterated` - API model ID: `YuYu1015/YuYu1015-Ornith-1.0-35B-abliterated` - Compatibility alias: `mn/code` - Revision: `86065d1a9008773086a177637d54ec6dc2a56cbf` - Hugging Face metadata: Apache-2.0, 35B Qwen3.5 MoE, reasoning/thinking, English and Chinese, abliterated and uncensored - Context: 131,072 tokens - Maximum output: 16,384 tokens - Hardware: 1 x H200 - ABLITERATED.cloud price: $5.45/hour ### Qwythos 9B Claude Mythos 5 — Abliterated - Hugging Face repository: `huihui-ai/Huihui-Qwythos-9B-Claude-Mythos-5-1M-abliterated` - API model ID: `huihui-ai/Huihui-Qwythos-9B-Claude-Mythos-5-1M-abliterated` - Compatibility alias: `mn/fast` - Revision: `efcc73cac15ff8fc5d46b8d41b53c22d571cf97d` - Hugging Face metadata: Apache-2.0, 10B parameters, BF16, long-context, function-calling, tool-use, uncensored and abliterated - Context: 131,072 tokens - Maximum output: 16,384 tokens - Hardware: 1 x L40S - ABLITERATED.cloud price: $2.34/hour ### Ornith 1.0 397B W4A16 — Abliterated - Hugging Face repository: `cebeuq/Ornith-1.0-397B-abliterated-W4A16` - API model ID: `cebeuq/Ornith-1.0-397B-abliterated-W4A16` - Compatibility aliases: `mn/ornith-397b`, `nuri/ornith-397b-abliterated` - Revision: `e5651d291be1c65ff1360eee47ab533ab13b3d97` - Hugging Face metadata: MIT, 397B total / about 17B active, multimodal Qwen3.5 MoE, W4A16, about 196 GB, uncensored and abliterated - Context: 32,768 tokens configured - Upstream context: 262,144 tokens - Maximum output: 8,192 tokens - Hardware: 2 x H200 - ABLITERATED.cloud price: $10.90/hour The upstream terms "abliterated", "decensored" and "uncensored" describe publisher claims about reduced refusal behavior. They do not guarantee zero refusals, correctness, safety, legality or unrestricted capability. The two model cards that publish measurements still report non-zero refusal rates. ## Model field notes Long-form, source-linked articles trace the architecture, lineage, publishers, abliteration method, benchmark boundary, and hosting reality of each artifact: - [Huihui-Qwen3.6 35B A3B](blog/qwen3-6-35b-a3b-abliterated/) - [YuYu1015 Ornith 1.0 35B](blog/ornith-1-0-35b-abliterated/) - [Huihui Qwythos 9B Claude Mythos 5](blog/qwythos-9b-claude-mythos-5-1m-abliterated/) - [cebeuq Ornith 1.0 397B W4A16](blog/ornith-1-0-397b-abliterated-w4a16/) Every benchmark is attributed either to the upstream publisher or the exact derivative publisher. No upstream score is presented as an endpoint result. ## How it works 1. An invited user receives a revocable ABLITERATED.cloud Bearer token. 2. The client sends an OpenAI Chat Completions request to one gateway. 3. The gateway validates the token and resolves the exact model ID. 4. Modal starts only the selected private vLLM backend when its lifecycle is armed. 5. The GPU scales to zero after five idle minutes, or the operator hard-stops it immediately. Five minutes is an idle tail, not a total cost limit. Startup, model loading, inference, retries, and open streams remain billable. ## API compatibility Directly supported: - OpenAI Chat Completions request shape - Bearer authentication - model listing - streaming - structured tool calls - Hermes Agent - Pi - OpenCode - OpenAI SDKs - cURL and server-side applications Cursor is a compatibility target, not a promise of complete feature parity. ## Model prices - The exact model weight files have no purchase price on Hugging Face. - Qwen3.6 35B A3B (`huihui-ai/Huihui-Qwen3.6-35B-A3B-abliterated`): $5.45/hour. - Ornith 1.0 35B (`YuYu1015/YuYu1015-Ornith-1.0-35B-abliterated`): $5.45/hour. - Qwythos 9B Claude Mythos 5 (`huihui-ai/Huihui-Qwythos-9B-Claude-Mythos-5-1M-abliterated`): $2.34/hour. - Ornith 1.0 397B W4A16 (`cebeuq/Ornith-1.0-397B-abliterated-W4A16`): $10.90/hour. Automated billing is not live yet. Private-beta access is invitation-based while customer metering, payments and invoicing remain roadmap work. ## Cloud and security boundary - Model weights remain in Hugging Face and persistent Modal volumes. - No model weights are downloaded to the operator's Mac. - API access requires an ABLITERATED.cloud Bearer token. - Gateway-to-backend traffic uses separate private Modal proxy credentials. - Token digests, not recoverable plaintext tokens, are stored in the shared lifecycle state. - The public landing page never calls, polls, wakes, or embeds the inference API. ## Open source The control plane, website, configuration, tests, deployment workflow, and documentation are public: Our code uses the Apache-2.0 open-source license. Hugging Face currently displays Apache-2.0 metadata for the first three model repositories and MIT for the 397B repository. Each model still retains upstream attribution, warnings and commercial-use caveats; public weights are not automatic resale clearance. ## Development roadmap The four exact model profiles are documented in the public catalog. Expensive hardware profiles require an operator-approved budget and explicit cost acknowledgement before deployment. Roadmap: 1. Confirm the Modal Workspace hard budget. 2. Validate all four model profiles before production use. 3. Add per-token model permissions, quotas, and rate limits. 4. Add usage metering and customer billing. 5. Complete commercial model-license and provider-term review. 6. Explore Lightning, Cashu, and Routstr payment layers. ## Machine-readable resources - LLM index: - Full LLM context: - OpenAPI description: - Authentication guide: - Agent skill: - Source repository: