Fine-tuned Qwen3-27B 'Humanlike Chat' passes as human in blind tests
solyarisoftware · x · 2026-10-12
A developer fine-tuned Qwen3.8-27B into 'Humanlike-Chat 2.0', aiming to make a local model chat like a real person instead of an AI assistant.
- In the creator's blind test, official Qwen3.8-27B was judged human 15.1% of the time vs 23.5% for Humanlike 2.0
- Trained on real + synthetic conversations, then on-policy distilled — no 'talk like a human' system prompt needed
- Retains tool calling, code, instruction following, reasoning, optional vision, 262K context
- Local quant is just 15.1GB (IQ4XS); runs on llama.cpp / LM Studio / Ollama; Apache 2.0
More from Models
- Gemini 4 Argon pricing ($2/$10) spotted on Google's SKU page, global release imminent — gaganghotra_ · 2026-10-13
- A style guide for making Claude-like models write normally, and its linguistic traps — gleech · 2026-10-13
- OpenAI denies Pro 200 grandfathering to a subscriber billed through the eligibility window — IrrelevantNecessity · 2026-10-13
- autotrust's JEV-35B-NVFP4, a Qwen3.5 MoE-based decision model, hits HF trending — autotrust · 2026-10-13
- Theo benchmarks Opus 5.5, GPT-6.1 Sol and Muse Wonder Full on a bug fix — shadcn · 2026-10-13
- Claude Code at $100/month plus $20 ChatGPT: AI subscription costs stack up — dumpshoot · 2026-10-13