AI Content Humanizer — Make AI Text Sound Human (Mistral + Nvidia NIM + blader/humanizer)

Transform robotic AI-generated content into natural, engaging human-like text. Built for the 2026 reality: Mistral free-tier limits + reliable Nvidia NIM fallback + the blader/humanizer 35-pattern method.

What it does

AI Content Humanizer converts machine-generated text into natural, human-sounding content via a 2-layer resilient pipeline:

1. Primary: Nvidia NIM — openai/gpt-oss-20b (default, free endpoint Available), nvidia/nemotron-3-nano-omni-30b-a3b-reasoning (fallback) via https://integrate.api.nvidia.com/v1 (OpenAI-compatible, NVIDIA_API_KEY)
2. Backup: Mistral — mistral-small-latest, open-mistral-7b, mistral-tiny, open-mixtral-8x7b, mistral-medium-latest, mistral-large-latest

Nvidia NIM is primary because Mistral's free tier is heavily rate-limited (429s). If the Nvidia chain fails, the actor falls back to Mistral (with retry + backoff); an explicit Mistral model reverses the order. You can also select a Mistral model directly.

Verified 2026-09-04: NIM lists 81 models, but only openai/gpt-oss-20b (free endpoint Available) and nvidia/nemotron-3-nano-omni-30b-a3b-reasoning are entitled for a standard key — others (llama-nemotron, kimi, palmyra, mistral-large-2, ...) return 404 "Not found for account". openai/gpt-oss-120b went end-of-life 2026-09-03 (free endpoint Deprecated, API returns 410 Gone) and was removed.

Why

• No more 403s: free-tier safe default mistral-small-latest + automatic fallback chain
• Bypass AI detectors via the blader/humanizer method: 35 patterns from Wikipedia's "Signs of AI writing" (inflated claims, sales language, stock AI words, em-dash abuse, chatbot artifacts, filler) with a two-pass draft-and-check rewrite
• Match your voice via optional voiceSample — provide 2-3 paragraphs of your writing and the rewrite follows its rhythm and quirks
• Preserve meaning 100% — same facts, human voice, never invents names/numbers/dates/citations
• SEO / AIO / GEO ready — natural flow ranks better in search and AI answers

Input

{
  "content": "Machine learning algorithms utilize statistical methods...",
  "model": "openai/gpt-oss-20b",
  "nvidiaModel": "nvidia/nemotron-3-nano-omni-30b-a3b-reasoning"
}

Field	Required	Description
content	yes	AI text to humanize (also accepts contents[0] or text for legacy)
model	no	Primary model. Default openai/gpt-oss-20b (Nvidia NIM). Use a mistral-* id to lead with Mistral instead
nvidiaModel	no	Override Nvidia NIM model (openai/gpt-oss-20b or nvidia/nemotron-3-nano-omni-30b-a3b-reasoning)
voiceSample	no	2-3 paragraphs of your own writing for voice matching (max 4000 chars, takes priority over style rules)

All models available:

• Nvidia NIM: openai/gpt-oss-20b (default), nvidia/nemotron-3-nano-omni-30b-a3b-reasoning (fallback)
• Mistral (backup): mistral-small-latest, open-mistral-7b, mistral-tiny, open-mixtral-8x7b, mistral-medium-latest, mistral-large-latest

Output

Pushed to dataset:

{
  "humanized": "You know how ML algorithms are like that super observant friend...",
  "modelUsed": "mistral-small-latest",
  "provider": "mistral",
  "method": "blader-humanizer-2.11.2",
  "duration": 2.8,
  "originalLength": 191,
  "humanizedLength": 761
}

provider is mistral or nvidia. modelUsed is the actual model that succeeded (may differ from input if fallback occurred). method tracks the humanizer skill version embedded in the system prompt.

How it works

1. Input → truncate to 12k chars if needed, prepend voiceSample (max 4k chars) when provided
2. Humanize with the blader/humanizer system prompt (35 AI-writing patterns, two-pass draft-and-check, no invented facts, em-dash ban)
3. Try Nvidia NIM chain first (openai/gpt-oss-20b → nvidia/nemotron-3-nano-omni-30b-a3b-reasoning, retry with backoff on 429/5xx). On failure, fall back to Mistral chain (preferred → env MISTRAL_MODEL → free-tier chain, same retry policy)
4. Charge PAY_PER_EVENT api-call $0.005 only after success
5. Push to dataset

Explicit Mistral model reverses the order (Mistral first, Nvidia backup).

Environment Variables

Var	Required	Description
MISTRAL_API_KEY	At least one of Mistral/Nvidia	@MISTRAL_API_KEY secret
NVIDIA_API_KEY	At least one	@NVIDIA_API_KEY secret (nvapi-...)
NVIDIA_BASE_URL	no	Default https://integrate.api.nvidia.com/v1
MISTRAL_MODEL	no	Override default Mistral model
NVIDIA_MODEL	no	Override default Nvidia model

Free-tier tip: mistral-large-latest requires paid plan → 403 tier_not_allowed. Keep default mistral-small-latest or enable Nvidia fallback.

Local dev

apify secrets add MISTRAL_API_KEY <key>
apify secrets add NVIDIA_API_KEY <nvapi-key>
apify run --purge
# or: MISTRAL_API_KEY=xxx NVIDIA_API_KEY=nvapi-... node src/main.js

Tested: mistral-small-latest 98→750 (3s), mistral-large → fallback to small 191→761 (2.8s), openai/gpt-oss-120b 211→334 (2.2s), large 2471→3378 (8s).

Technical specs

• Engine: Node 20, apify@3.4.4, @mistralai/mistralai@2.1.2, native fetch for NIM
• Prompts: blader/humanizer v2.11.2 method (35 AI-writing patterns, two-pass rewrite, voice matching)
• Temperature: 0.7, top_p 0.9, max_tokens 8192
• Memory: 256 MB
• Pricing: $0.005 / successful humanization

Credits

• Humanization method: blader/humanizer v2.11.2 (MIT) — installed in this repo under .agents/skills/humanizer via npx skills add blader/humanizer. The actor embeds a condensed version of its 35 patterns in the system prompt (src/main.js).
• Pattern source: Wikipedia's "Signs of AI writing", maintained by WikiProject AI Cleanup.

Data security

• No retention after processing, GDPR compliant

---

Keywords: AI humanizer, humanize AI text, Mistral, Nvidia NIM, gpt-oss-20b, AI to human text, content humanization