dev-5090 uncensored inference

Uncensored general chat (Qwen3.8-27B Heretic).

[preview] Image understanding. Currently TEXT-FALLBACK ONLY — no vision model is loaded. Priced accordingly.

Text embeddings (768-dim). Send {"input": "text"} or {"input": ["a","b"]}.

Search a namespace you indexed: the query is embedded on the same model and matched against your collection. Returns hits with score, text and metadata, plus latency.

Fix broken JSON (trailing commas, single quotes, unquoted keys). Returns only valid JSON.

Explain a regex in plain English. Paste the regex in message content.

Convert a curl command to Python/JS/Go. Specify target language in the message.

Explain a SQL query in plain English. Paste query in message.

Explain code snippet in plain English (paste code in message).

Build a regex from a natural-language description. Returns just the regex + one test string.

Translate text into a target language (ISO code: es, de, fr, ja, pt, pl...). Optional domain and tone hints. Returns the translation, the model, usage and latency.

Uncensored coder completions, served by Qwen3.8-27B abliterated (Heretic) on a local RTX 5090 via vLLM.

Refactor code while preserving behavior. Returns diff + rationale.

Generate tests for a code snippet (matches file's existing test framework).

Multi-step reasoning and planning, served by Qwen3.8-27B abliterated (Heretic) on a local RTX 5090 via vLLM.

Structured code review: correctness, security, edge cases, reuse. Returns numbered findings.

Agent-readiness audit of any public URL: score /100 with per-check raw evidence, published weights, and a recompute recipe. Deterministic — no model, same input gives the same score.

Summarize a document without it leaving this machine. mode=bullets|paragraph|key-points, max_words caps the length. Returns the summary, the model that wrote it, token usage and latency.

Embed and index documents into a private vector collection (768-dim, cosine). documents=[{id,text,metadata}], namespace picks the collection. Returns documents, chunks, tokens and latency.

Security audit of code snippet (SQLi, XSS, auth bypass, secret leaks, unsafe deserialization).

Complete a full agentic task (plan → generate → verify). Longer running.

Proof-of-Model: detect silent model substitution at any OpenAI-compatible inference endpoint. mode=fingerprint records what a seller serves today; mode=verify re-probes later and proves whether the weights changed. Deterministic, evidence included, no model used to judge.

Pay-to-compose orchestrator: pulls live data (weather), routes it through two local-model stages, then an independent judge verifies the result against the source data before it ships (one bounded self-correction on rejection). Returns the composed answer, the judge's verdict, and a per-step cost ledger.