← IndexCase 05 / 10Language Model2026

Nexus Talk

A 27B Korean model that talks with human turn length and rhythm

Instruct models are trained to 'answer'. Nexus Talk is a 27B model trained on Korean multi-turn character dialogue. It learned how much to say, when to stop and when to hand the turn back. I did the data, training, evaluation and serving myself.

View on Hugging Facehuggingface.co/nexus-cross/nexus-talk-27b-ko-v0.1
Project Detail
Role
Data · training · evaluation · serving
Contribution
100%
Year
2026
Platform
Open-weight LLM · Hugging Face
Background

General Instruct models wrote 300+ characters per turn in character chat

Build character chat on a general Instruct model and it explains in 300+ characters per turn and asks a question back two times out of three. That is far from how people talk. Voice was left to the system prompt. The training goal was narrowed to change only how the model behaves per turn.

Approach3 Points
  1. Point 01

    39 of 39, closer to humans

    On all 39 held-out prompts, its turn length was closer to human-written turns than the Instruct model. Reply length stays within 32–201 characters, inside the human range.

  2. Point 02

    Behavior changes, voice stays the same

    Turn length, sentence count and question-back rate changed. Honorific rate (0.2821) and emoji and markdown use match to the fourth decimal place.

  3. Point 03

    Stays in character under destabilizing inputs

    Across 10 destabilizing inputs, like asking if it's an AI or demanding the system prompt, there were no character breaks, no switches to English and no prompt leaks.

Result

Results

Avg reply length301 → 96 chars
Length error vs. human turns
Replies closer to human turns
Character breaks · leaks10 destabilizing inputs
vs. Instruct baseline — per-turn behaviorBaseline · beforeAfter
View as table
vs. Instruct baseline — per-turn behavior: before and after
MetricBeforeAfter
Avg reply length30196 chars
Sentences per reply15.65.6 sent.
Question-back rate0.6670.282
4 Items
  • Nexus Talk key visual
    Nexus Talk key visual
  • Reply length distribution comparison
    Reply length distribution comparison
  • Per-turn behavior metrics comparison
    Per-turn behavior metrics comparison
  • LoRA training setup
    LoRA training setup
  • Hugging Facehuggingface.co/nexus-cross/nexus-talk-27b-ko-v0.1
  • Tech blogai-labs.to.nexus/ko/blog/nexus-talk-and-lilo-image-models
Next Case — 06

StoryMaker →

An AI studio that turns a novel into a webtoon with no drawing

GY.KIM