Frustration with overly polite, verbose, and uniform LLM responses may be the least celebrated driver of open source. From that irritation comes Qwen3.8-27B-Humanlike-Chat, a model put together by a creator known on Reddit as kvyb to imitate real human-to-human conversation. It is not a race for benchmarks: the stated goal is to change conversational habits, removing the tendency to explain everything, to agree with the interlocutor, and to fill silence just to keep the exchange going.

To do it, kvyb built a dataset of 125,217 obfuscated messages drawn from 1,396 human-to-human conversations and trained a rank-256 LoRA on top of Huihui-Qwen3.8-27B-abliterated, a variant of the 27-billion-parameter Qwen family. The released version is checkpoint 863. In informal testing, the result feels less like an assistant and more like a conversation partner: shorter, less polished, more natural replies, even without a system prompt.

The delicate point is the price paid on the reliability front. An earlier iteration scored five percentage points lower than the Huihui parent model on IFEval, an instruction-following benchmark. The creator has not rerun that measurement on checkpoint 863 and has not tested coding performance: a verification gap that matters for anyone evaluating production use. A five-point drop on IFEval may not be a detail: many automated workflows depend on precise instructions. But the project was not born to replace an enterprise assistant; it was born to reduce friction in informal conversations, where the help-desk tone is itself a bug.

GGUF format and the self-hosted door

Publishing merged GGUFs and the standalone F32 LoRA adapter lowers the threshold for anyone who wants to try the model locally, without going through a cloud API. From an on-premise perspective, the GGUF format is now a de facto standard for inference on consumer and workstation hardware, because it allows Quantization to reduce VRAM usage. The fact that the project also distributes a free rate-limited OpenAI-compatible endpoint, along with a demo on Hugging Face Space, shows the dual nature of these releases: immediate cloud access and the possibility of self-hosting.

The connection with AI-RADAR is nearly immediate. A model like this touches the core of the trade-off between control and convenience: anyone who wants a specific conversational tone can start from an open checkpoint, apply a LoRA adapter, and serve the model locally, thus keeping sovereignty over conversation data. But the choice is not free: the hardware needed for inference of a 27-billion-parameter model, even with Quantization, requires careful evaluation of VRAM and operating costs. For those evaluating on-premise deployment, there are trade-offs between model quality, latency, and TCO that cannot be resolved with a simple benchmark.

The LessThanThreeAI Hugging Face repository hosts the merged GGUFs and the F32 adapter; a Space lets users try the model with different system prompts and reasoning modes, while the rate-limited OpenAI-compatible endpoint offers immediate access. The choice to obfuscate the dataset messages adds another signal: even in a project born to make models more conversational, protecting the texts used for fine-tuning was not treated as an afterthought.