Qwen3.8-27B-Humanlike-Chat 2.0: texts like a human, now with tool calls and better instruction following

Last month I posted a Qwen3.8-27B LoRA that makes it talk like a person instead of an assistant. It got a lot more attention than I expected: 700+ upvotes, 248 comments and 44k downloads since.

I read every comment. People really don't like assistant speak, so its tone of voice resonated. The rest got roasted, very fairly:

incapable of producing more than a few words at a time.

single default personality which no amount of prompting can overcome

will not use tools, at all, whatsoever.

There needs to be a middle ground

They were right. The tool calls didn't actually work, and when people asked it to do something it would sometimes just say it's busy or going to bed. Very human. In a bad way.

So I spent the last three weeks on 2.0. The goal was simple: keep the voice people liked and lose the drawbacks.

What 2.0 does now

• With no system prompt, it's a normal person texting. Not an assistant, not a catgirl.
• Give it a character card and it becomes that person, and still texts like one.
• Ask for a formal email, numbered steps or a proper explanation, and you get exactly that. Then it goes back to texting.
• Don't want the lowercase texting? Tell it "from now on write in full sentences" (or put it in the system prompt) and it sticks to that until you say otherwise. v1 ignored this completely.
• It calls tools, and it asks when something is missing instead of making it up. This is the part I'm happiest about. Ask the base model to book a flight without saying where from and it picks JFK. 2.0 asks where you're flying from.
• It writes code and does math at roughly base-model level.

It's a colleague and a humanlike companion, not an assistant. Use it for chat, roleplay, agents or actual work.

How I trained it

v1 was plain SFT on real and synthetic conversations (139,845 messages from 1,396 conversations). That copies habits, including the bad ones.

For 2.0 I used on-policy distillation. The model writes its own replies and a teacher grades every token. There are two teachers:

• v1 plus a hidden "text like a person" instruction, for chat and characters;
• the plain base model, for instructions, tools and code.

The student never sees the hidden instruction, so it learns the behaviour without needing a prompt. Same 27B, a second LoRA on top, merged.

Numbers (vs the model I trained on, huihui-ai's abliterated Qwen3.8-27B; same prompts, same run, thinking off)

Benchmark	Base (abliterated)	2.0
IFBench (instruction types I never trained on)	37.3	43.7
When2Call (call, ask or refuse correctly)	48	58
BFCL irrelevance (don't call a tool when none fits)	60	78
IFEval, GSM8K, BFCL simple	81.9 / 89.1 / 97	83.5 / 89.1 / 98 (ties)

Full chart in the images.

Where it's still worse: knowledge (MMLU-Pro 72.5 vs 78.5) and competitive code (LiveCodeBench 51 vs 56).

Is it actually more human? I built a benchmark for this, "ishuman":

• It takes 150 fragments from unseen chats.
• Has each model write the next message.
• Shows a judge the real message and the model's without labels, and asks which one a person wrote.

Model	Judge thought it was the real person (50% = can't tell)
Qwen3.8-27B abliterated (huihui-ai, the model I trained on)	0.3%
Same abliterated model + a "text like a human" system prompt	6.8%
Qwen3.8-27B official (unmodified, via OpenRouter)	15.1%
Qwen3.8-27B-Humanlike-Chat 2.0	23.5%

So no, you can't just prompt your way there. In a separate test of 16 live multi-turn chats with invented people, 2.0 was picked over the base model 16 out of 16 times.

Links

• Hugging Face: GGUFs (IQ4_XS, Q4_K_M, Q5_K_M, Q6_K, Q8_0, BF16) plus the standalone LoRA: https://huggingface.co/LessThanThreeAI/Qwen3.8-27B-Humanlike-Chat-GGUF
• Demo space: https://huggingface.co/spaces/LessThanThreeAI/Qwen3.8-27B-Humanlike-Chat

Big thanks to everyone who left feedback last time, especially the ones who were critical. Tell me where it still sounds like an assistant.