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.

Pico

≤100M parameters

Micro

300M – 2B parameters

Mini

4B – 6B parameters

Small

8B – 14B parameters

Medium

16B – 50B parameters

Large

60B – 200B parameters

The size word in every model's name — Pico, Micro, Mini, Small, Medium, Large — is a fixed parameter-count band, the same across every family below. The gaps between tiers (100M–300M, 2B–4B, 6B–8B, 14B–16B, 50B–60B, and above 200B) are simply ranges we haven't shipped a model in yet.

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.

Model	Base	Size	Result
Prism Caption 1 Micro Chat titling	Gemma-3-1B	1B · Micro	0/20 bad-format outputs (base: 3/20)
Prism Caption 1.5 Micro Chat titling	Qwen3-0.6B	0.6B · Micro	0/24 format issues, 23/24 within spec
Prism Caption 2 Micro Chat titling	LFM2-700M	742M · Micro	0/275 formatting issues, 273/275 relevant across three eval rounds at increasing scale (24 → 145 → 275 held-out topics)
Prism Caption 2.5 Micro Chat titling	LFM2-350M	354M · Micro	0/275 formatting issues, 273/275 relevant -- matches the 700M Prism Caption 2 Micro at less than half the parameters
Prism Caption 3 Micro Chat titling	LFM2-350M	354M · Micro	0/275 formatting issues, 275/275 relevant at 354M parameters; ~63 ms per title
Prism Caption 3 Pico Chat titling	SmolLM2-135M-Instruct	135M	2/275 formatting issues, 275/275 relevant at 135M parameters; ~48 ms per title
Prism Creative 1 Mini Creative writing	Qwen3-4B	4B · Mini	Coherent, in-genre output with zero leaked planning text or repetition loops
Prism Creative 1.5 Mini Writing, editing, and revision	Qwen3-4B	4B · Mini	3.2x lower repetition than base; consistent constraint-following on tightening and rewrite tasks
Prism Roleplay 1 Small Character roleplay	Qwen3-8B	8B · Small	100% suppression of leaked reasoning text (base: 100% leaked); 3x lower repetition
Prism Roleplay 1.5 Small Character roleplay, quality-focused	Qwen3-8B	8B · Small	83% blind-judge win rate vs. base (10W–2L–4T across 16 varied scenes)
Prism Roleplay 1.5 Small Fast Fast, small-footprint character roleplay	Bonsai-8B (1-bit)	8B (mixed 1-bit / 6-bit)	Beats base Qwen3-8B in 78% of blind comparisons at 28 tok/s in 3.6 GB (1.3x faster, 26% less memory than the 4-bit base); Prism Roleplay 1.5 Small still wins head-to-head 59% of the time
Prism Safety 1 Micro Content-safety classifier	Qwen3-1.7B	1.7B · Micro	78.4% accuracy, 86.5% recall — vs. 84.4% / 90.8% for a 5x larger frontier guard model
Prism Creative 2 MiniTraining Writing specialist, next iteration	Qwen3-4B	4B · Mini	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.

Model	Base	Size	Result
Amethyst 1 Mini General chat	Gemma-3-4B	4B · Mini	First-generation validation of the distillation pipeline
Amethyst 1 Small General chat	Llama-3.1-8B	8B · Small	Same training data, larger base model
Amethyst 1.5 Mini Chat + web-search tool-calling	Gemma-3-4B	4B · Mini	16/16 tool-call correctness on held-out prompts (base: 5/16)
Amethyst 2 Mini (Failed) General chat with reliable web-search tool calling	Qwen3-4B	4B · Mini	Decides correctly whether to call web_search on 28/28 held-out prompts (base Qwen3-4B: 24/28) and wins general chat 19-9 in blind pairwise judging

Copal

Agentic tool use

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

Model	Base	Size	Result
Copal 1 Mini Agentic tool-calling	Gemma-3-4B	4B · Mini	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.

Model	Base	Size	Result
AquamarineTraining Code generation, debugging, refactoring	Qwen3-4B-Instruct-2507	4B · Mini	In active data generation and training

Quartz

From-scratch pretrain

Not a LoRA distillation like the others — randomly-initialized weights, a DeepSeek-style mixture-of-experts trained from raw token data on a single consumer GPU.

Model	Base	Size	Result
Quartz Micro Preview Full-parameter instruction tuning of the from-scratch pretrain	Quartz Micro Preview Base	1.03B · Micro	Full-parameter SFT on databricks-dolly-15k, 1,317 steps. Modest, mixed improvement over the base checkpoint: clearly better at grounded/context QA, still weak at strict format- and length-following.
Quartz Micro Preview Base From-scratch pretrain (base checkpoint, pre-instruction-tuning)	Randomly-initialized weights	1.03B · Micro	24,414 steps on 100M tokens. Raw base model -- no instruction-following behavior yet; that's what the upcoming instruction-tuned release is for.

CodeIn

One language, one specialist

Where Aquamarine is a generalist across languages, CodeIn models go the other direction: each one distilled for exactly one programming language, and nothing else.

Model	Base	Size	Result
CodeInLuau 1 Small Luau code completion	Qwen3-8B	8B · Small	18% real syntax-error rate on base vs 2% tuned, on 100 held-out real-corpus prompts (luau-analyze)
CodeInSwift 1 Small Swift coding specialist	Qwen3-8B	8B · Small	38% of held-out prompts produce working, swiftc-typechecked code within a 512-token budget (base: 0%, exhausted by unbounded chain-of-thought before reaching any code)

Cybermarine

Cybersecurity specialist

Aquamarine's coding skills pointed at security: agentic software-engineering trajectories, defensive security-agent behavior, and CVE/CWE vulnerability knowledge, built on our CyberData set.

Model	Base	Size	Result
Cybermarine 1 SmallPlanned Cybersecurity and agentic coding	Qwen3-8B (Heretic)	8B · Small	Dataset ready (CyberData Small, 15,000 examples; Medium, Large and Full also available); training hasn't started yet