Only implemented or actively validating changes appear here. Planned capability remains on the roadmap until evidence exists.
2026-09-23
24 September 2026 · Site freshness restore
2026-09-24 (SGT) — Site freshness restore: text inference marketed on RTX 5090 with 131,072-token context; vision/multimodal copy restored (Huihui VL sidecar via image_url); Bitcoin checkout open (beta), card closed; partnership promo UI restored on Account. Admin gate cache remains ?v=20260924admin5.
Vision / multimodal Chat Completions
Image chat is available on the same API. Requests with image_url (data-URL only) route to Huihui-Qwen3-VL-8B-Instruct-abliterated. Text traffic stays on qwen3.8-27b-abliterated / RTX 5090. Remote http(s) image URLs and audio remain unsupported.
5 September 2026 — release candidate, not yet deployed
Prepared a monochrome dimensional website, practical account and billing guides, separate live readiness checks, readable reference pricing, and refreshed client support disclosures. Public Bitcoin checkout via BTCPay is available; card closed. These changes describe an undeployed release candidate until a separately approved release.
4 September 2026 · Release A deployment and validation
Deployed corrected public pack values while keeping every payment route fail-closed.
Published exact launch hardware, checkpoint revision, quantization, context, and unsupported modalities.
Activated the live catalog contract and readiness endpoint.
Confirmed the running model process, local snapshot, and repository ref at revision c2f5341e7c9a1c55e6d61cfc8e6d0ca897bd7443, with vLLM 0.28.0, 8,192 total context, two sequences, and non-thinking defaults.
Added first-success API documentation, integration gates, structured metadata, FAQ, model disclosure, and privacy engineering note.
Passed local model, text, streaming, error, and accounting loops with pinned OpenAI Python 3.8.0 and Node.js 7.10.0 clients.
Recorded local/mock sequential and parallel tool protocol fixtures with pinned OpenCode, Codex, and unofficial Claude Code clients; standard tool-bearing requests remain unsupported until exact-GPU parser validation.
Published non-secret, non-overwriting reference-only configuration downloads and setup pages for all three coding harnesses.
Release policy
Production dates are added after deployment and smoke tests, not before. Changes to the model, quantization, context, parser, or serving stack restart the relevant acceptance clock.