# ABLITERATED.cloud website releases ## website-v0.12.4 [Exact numbers in plain-language facts](releases/website-v0.12.4.md). ## website-v0.12.3 [Plain-language facts and article SEO layer](releases/website-v0.12.3.md). ## website-v0.12.2 [Intelligence, freed. and faster-reading model research](releases/website-v0.12.2.md). Restored cloud-GPU, self-hosting and LLM-router positioning; direct blog access; source-backed TL;DR and Basically facts throughout the article inventory. ## website-v0.12.0 [Vast-only documentation, MIT and lightweight website](releases/website-v0.12.0.md). See the repository changelog for operating-status corrections and the archive map. Publication is recorded by the signed GitHub Pages release. ## Historical website releases The entries below describe past releases, not current graphics, service availability, prices or licensing of this release. ## website-v0.8.2 Pastel synapses now span the complete hero instead of forming a brain. ### Changed - Fill the hero with the deterministic 1,280-node Canvas network on desktop and mobile. - Use pastel pink, blue, green, peach and violet for resting connections and pointer-activated glow. - Remove the brain outline, cortical folds and static image fallback. - Preserve copy contrast, pointer interaction, viewport pausing and reduced motion behavior. ### Safety and operations - Do not deploy, wake or start any Modal model or GPU. - Preserve all existing model runtime settings. ## website-v0.8.1 The hero brain now carries substantially more synaptic connections while its left side fades gently beneath the copy. ### Changed - Raise the maximum local connections per point from six to ten. - Widen the local connection radius from `0.145` to `0.165`. - Restore a directional background-colored overlay on the left side only, preserving a crisp, fully visible right hemisphere. - Keep the Canvas dependency-free, locally generated and pointer-reactive. ### Safety and operations - Do not deploy, wake or start any Modal model or GPU. - Preserve the deployment-disabled Ornith 397B profile. ## website-v0.8.0 Four model stories, a denser hero brain and a complete brain-shaped identity. ### Model field notes - Publish one deeply researched, source-linked article for every exact model artifact in the catalog, in both semantic HTML and Markdown. - Explain architecture, lineage, training, abliteration, quantization, published evaluations, hosting implications and the publicly verifiable publisher history without inventing a model crew. - Keep publisher measurements labeled and never transfer an upstream benchmark to an abliterated derivative that has not been rerun. - Add Blog and TechArticle structured data, canonical URLs, RSS, sitemap and agent-readable discovery through `llms.txt` and `llms-full.txt`. - Preserve canonical trailing slashes through the prepared edge Worker so blog directory requests do not enter a redirect loop. - Replace informational OAuth endpoint stubs with a functional anonymous `client_credentials` flow that grants only `public:read` metadata access. ### Visual identity - Increase the interactive hero to 1,280 deterministic nodes, six connections per point and eight visible cortical fold paths. - Remove the hero fade so the complete brain and its synapses remain visible. - Use a connected brain mark for the navigation logo, SVG favicon, PNG device icons and legacy ICO favicon. - Change the supporting line to “Abliterated open models. One simple API.” ### Safety and operations - Do not deploy, wake or start any Modal model or GPU. - Preserve the deployment-disabled two-H200 Ornith 397B profile. - Extend static and local edge checks for blog integrity and icon completeness. ## website-v0.7.1 The production website now uses `https://abliterated.cloud/` as its GitHub Pages custom domain and canonical origin. ### Custom domain - Replace Porkbun parking records with GitHub's official apex A and AAAA targets plus the direct `www` CNAME. - Add the matching `website/CNAME` file and repository Pages setting. - Update canonical, social, structured-data, sitemap, robots, LLM index, and Agent Skill URLs to the production origin. ### Operations - Document GitHub TLS provisioning and exact public verification commands. - Keep Cloudflare as an optional prepared edge target; this release deploys only static GitHub Pages and does not start Modal or any GPU. ## website-v0.7.0 Exact model identities everywhere, a denser background brain, and cleaner public product copy. ### Changed - Make the four exact Hugging Face `owner/repository` values the primary API model IDs across the page, OpenAPI, agent documents and integrations. - Keep the previous `mn/*` names only as compatibility aliases. - Replace role labels such as `god`, `code` and `fast` with real model-family names and recognizable repository-derived Modal app slugs. - Change the hero to “Intelligence, freed. Uncensored, abliterated AI.” and introduce the product as “Abliterated open models in the cloud. One simple API.” - Move the brain behind the hero copy, increase it to 820 deterministic nodes, and remove its instructional label. - Remove internal endpoint lifecycle labels from the public landing page and machine-readable catalog. ### Runtime safety - Prepare the renamed runtime definitions without deploying them or starting a GPU. - Keep the two-H200 Ornith 397B profile deployment-disabled until a signed, budget-approved runtime release explicitly enables it. ## website-v0.6.0 A recognizable interactive brain and a more explicit hero statement. ### Changed - Replace the loose point-cloud volume with a closed, smoothed lateral brain silhouette, including the cerebrum, lower rear form and small brain stem. - Fill the silhouette with 420 deterministic connected nodes and retain a stronger contour so the shape stays legible at desktop and mobile sizes. - Light up local neural regions in cyan, magenta, lime, orange and violet as the pointer moves over their points and connecting lines. - Change the headline to “Intelligence, freed: uncensored, private abliterated AI.” using responsive line lengths. ### Performance and accessibility - Keep the interaction inside the existing dependency-free Canvas runtime. - Preserve offscreen and hidden-tab pausing, roughly 30 FPS throttling, capped device pixel ratio and static reduced-motion rendering with pointer feedback. ## website-v0.5.1 A typography refinement using one mono family and a strict 17 px minimum. ### Changed - Use a single Geist Mono-compatible local font stack across body copy, headings, navigation, buttons, model metadata, FAQ and code. - Increase every fixed font size below 17 px to 17 px, including status lines, repository labels, facts, hero annotation and footer links. - Adjust header height, controls, model-fact width and mobile brain annotation to keep the larger interface type readable without reintroducing clutter. ## website-v0.5.0 A brighter, narrower landing page with the clarity of a developer product, the complete four-model catalog, and a lightweight animated brain. ### Changed - Rebuilt the hero as one focused column followed by a wide point-cloud brain. - Tightened the page to a 1040 px content grid, flatter borders, shorter sections, compact mono metadata, and large readable sans-serif headings. - Added a visible **Star on GitHub** action while keeping Signal as the primary private-beta access path. - Preserved every exact Hugging Face repository, model status, context, hardware, license, API ID, price, caveat, FAQ and integration detail. ### Performance and accessibility - Implemented the brain with local Canvas code and no Three.js, CDN, webfont, analytics or other runtime dependency. - Pause animation when it leaves the viewport or the tab is hidden, render at no more than roughly 30 frames per second, cap device pixel ratio at 2, and show a static frame for reduced-motion users. - Retained the optimized owner-supplied brain artwork as the no-JavaScript and social-card fallback without preloading it on normal visits. ## website-v0.4.1 Fixed WebMCP discovery by registering the three public read-only catalog tools with `navigator.modelContext` when a supporting browser or scanner exposes it. The page remains dependency-free and never calls the inference API. ## website-v0.4.0 A quieter, larger-type landing page with readable model names, complete Hugging Face repository identities, and the Agent-Ready Cloudflare edge layer. ### Changed - Removed decorative API-window, badge, card and diagram treatments. - Reorganized the page into a short explanation, three steps, four model rows, one API example, transparency notes, FAQ and access request. - Made readable factual model names the headings, exact Hugging Face `owner/repository` paths the linked source field, and `mn/*` values the secondary API model IDs. - Increased body, model, section and FAQ typography and whitespace. ### Agent readiness - Added Markdown negotiation, discovery headers and exact media types through a Cloudflare Worker. - Added OAuth, A2A, MCP, Agent Skills, WebMCP and HTTP-signature discovery. - Added a local verifier for all application-controlled scanner requirements. - Kept GitHub Pages as the public preview until the final domain moves from Porkbun DNS to an active Cloudflare zone with DNSSEC and DNS-AID. ## website-v0.3.0 A calmer, more powerful identity centered on “Intelligence, freed.” ### Changed - Added the owner-supplied brain-and-broken-chains artwork as the responsive hero background, with optimized AVIF and WebP variants. - Replaced the previous hero statement with “Intelligence, freed.” and the original prompt line “Ask anything. Think for yourself.” - Simplified the palette, cards, borders, shadows, gradients, buttons, pricing, and closing section for a drier, more minimal interface. - Added clearer background about why the service exists: exact checkpoint identity, reduced refusal behavior, operator-controlled lifecycle, public prices, and inspectable source. - Changed every active access request from WhatsApp to Signal. - Simplified public pricing to the four ABLITERATED.cloud prices without publishing markup or margin mechanics. - Updated active Markdown, LLM, authentication, OpenAPI, social metadata, and discovery-facing documents to match. ### Current prices - Qwen3.6 35B A3B: $5.45/hour - Ornith 35B: $5.45/hour - Qwythos 9B: $2.34/hour - Ornith 397B W4A16: $10.90/hour Automated billing is not live yet. The 397B route remains prepared for the next budgeted runtime release. ## website-v0.2.0 Clearer model identity, free-weight explanation, SEO copy, and transparent planned pricing. ### Changed - Replaced shortened marketing names on all model cards with the complete real Hugging Face repository names. - Added source-backed parameter counts, architectures, upstream context, licenses, multimodal/reasoning/tool-use metadata, and exact pinned-model links. - Explained the differences between abliterated, uncensored, decensored, and the broader informal term jailbroken without promising zero refusals. - Made the free/paid boundary explicit: weight files are publicly downloadable; managed cloud GPU inference costs money. - Added per-model planned private-beta reference rates calculated as current Modal GPU list price multiplied by exactly 1.20. - Kept the base GPU cost beside every planned rate and disclosed that billing, metering, payments, and invoicing are not implemented yet. - Updated search, Open Graph, Twitter, JSON-LD, Markdown, LLM, FAQ, and social image copy. ### Planned reference rates - `huihui-ai/Huihui-Qwen3.6-35B-A3B-abliterated`: $5.44752/hour - `YuYu1015/YuYu1015-Ornith-1.0-35B-abliterated`: $5.44752/hour - `huihui-ai/Huihui-Qwythos-9B-Claude-Mythos-5-1M-abliterated`: $2.34144/hour - `cebeuq/Ornith-1.0-397B-abliterated-W4A16`: $10.89504/hour These are not live customer charges. ## website-v0.1.0 Initial open-source GitHub Pages release. ### Added - Original responsive ABLITERATED.cloud landing page inspired by the calm single-page product rhythm of receive.link without copying its GPL source, assets, fonts, icons, or text. - Exact four-model catalog with Hugging Face links, status, context, hardware, and transparent Modal base GPU estimates. - WhatsApp access requests through `wa.me/13103408213`. - Dependency-free HTML, CSS, and a small deferred navigation script. - Open Graph image, favicon, manifest, canonical URL, structured JSON-LD, semantic headings, FAQ markup, sitemap, and crawler policy. - Agent resources: `index.md`, `llms.txt`, `llms-full.txt`, `auth.md`, OpenAPI, extensionless API catalog, security contact, and versioned and legacy Agent Skill discovery indexes. - Signed branch-based GitHub Pages deployment isolated from Modal and model-release workflows. - Automated checks for metadata, structured data, internal resources, WhatsApp links, secret safety, runtime dependencies, and payload budgets. ### Verified before deployment - 60 repository tests passed. - Secret scan passed. - JSON, XML, YAML, and internal-resource validation passed. - Desktop and responsive mobile renderings were visually inspected. - Initial HTML, CSS, JavaScript, logo, and favicon total less than 125 KB. - No analytics, third-party runtime resources, webfonts, API polling, or Modal calls are loaded by the homepage. - The Pages deployment was built successfully from a signed `gh-pages` commit. - The custom Actions workflow was removed after GitHub refused to start it due to an account billing lock; the verified classic Pages build is used instead. ### Domain The initial deployment uses `https://eminogrande.github.io/mn-uncensored/`. The intended later custom domain is `https://abliterated.cloud/` after GitHub verification, Porkbun DNS cutover, HTTPS provisioning, canonical URL migration, and live SEO validation.