---

license: apache-2.0

datasets:

• openbmb/Ultra-FineWeb
• openbmb/Ultra-FineWeb-L3

language:

• en

pipeline_tag: text-generation

tags:

• chat
• tiny
• vertex

base_model: VertexResearch/Vertex-0.6-15M-Instruct

base_model_relation: quantized

---

Vertex 0.6 15M Instruct — GGUF

A tiny (~15M-param) chat model from the Vertex 0.6

family. Qwen3 architecture (hidden 256, 10 layers, GQA, 20000 vocab, ctx

2048), pretrained from scratch on 12B tokens of Ultra-FineWeb and

Ultra-FineWeb-L3 English (natural + synthetic-rewrite mix), then taken through a custom post-training pipeline for chat.

What it does: coherent multi-turn chat with in-context memory (recalls

your name/details from 1000+ tokens back, and says so honestly when you

haven't told it).

What it doesn't do: facts, reasoning, math, code. At this size,

knowledge is decorative — treat outputs as conversation, not information.

Usage

ChatML template (embedded). Sampling strongly recommended — greedy

decoding loops badly at this size:

temperature 0.6, top_p 0.9, repeat_penalty 1.3

GGUF builds:

Vertex-0.6-15M-Instruct-GGUF.

Trained end-to-end on a single RTX 4060 Laptop (8GB).

Training data

Fine-tuned from Vertex-0.6-15M-Base (pretrained on openbmb/Ultra-FineWeb + openbmb/Ultra-FineWeb-L3, 12B tokens). Instruction tuning was done with a custom post-training pipeline.