Models — Vertex AGI

Models

Every model we've shipped.

Grouped by family. Every result below comes from a held-out evaluation set with verified zero overlap against training data.

Prism

Single-purpose specialists

Narrow models distilled to do exactly one job — titling, safety classification, creative writing, roleplay — well enough to replace a frontier-model call.

Prism Caption 1 Micro

Gemma-3-1B · Chat titling

0/20 bad-format outputs (base: 3/20)

Prism Caption 1.5 Micro

0.6B

Qwen3-0.6B · Chat titling

0/24 format issues, 23/24 within spec

Prism Caption 2 Micro

742M

LFM2-700M · Chat titling

0/275 formatting issues, 273/275 relevant across three eval rounds at increasing scale (24 → 145 → 275 held-out topics)

Prism Creative 1 Mini

Qwen3-4B · Creative writing

Coherent, in-genre output with zero leaked planning text or repetition loops

Prism Creative 1.5 Mini

Qwen3-4B · Writing, editing, and revision

3.2x lower repetition than base; consistent constraint-following on tightening and rewrite tasks

Prism Roleplay 1 Small

Qwen3-8B · Character roleplay

100% suppression of leaked reasoning text (base: 100% leaked); 3x lower repetition

Prism Roleplay 1.5 Small

Qwen3-8B · Character roleplay, quality-focused

83% blind-judge win rate vs. base (10W–2L–4T across 16 varied scenes)

Prism Safety 1 Micro

1.7B

Qwen3-1.7B · Content-safety classifier

78.4% accuracy, 86.5% recall — vs. 84.4% / 90.8% for a 5x larger frontier guard model

Prism Creative 2 Mini

Training

Qwen3-4B · 4B · Writing specialist, next iteration

In data generation now — distilled from a rotating multi-teacher mix

Amethyst

General-purpose chat

Broader conversational models, including tool-calling variants, built for everyday assistant use rather than one narrow task.

Amethyst 1 Mini

Gemma-3-4B · General chat

First-generation validation of the distillation pipeline

Amethyst 1 Small

Llama-3.1-8B · General chat

Same training data, larger base model

Amethyst 1.5 Mini

Gemma-3-4B · Chat + web-search tool-calling

16/16 tool-call correctness on held-out prompts (base: 5/16)

Copal

Agentic tool use

Models trained specifically to decide when and how to call a tool — not just chat, but act.

Copal 1 Mini

Gemma-3-4B · Agentic tool-calling

11/12 appropriate tool use on held-out tasks (base: 8/12), zero parse errors

Aquamarine

Code specialist

A coding-distillation model in active training: generation, debugging, and refactoring across Python, JavaScript, TypeScript, Go, and SQL.

Aquamarine

Training

Qwen3-4B-Instruct-2507 · 4B · Code generation, debugging, refactoring

In active data generation and training