Small models.
Real evaluations.
Vertex AGI builds single-purpose language models distilled to do one job well, evaluated exclusively on held-out data — plus Quartz Micro Preview, a from-scratch pretrain, in the open.
Models shipped
Model families
Models in active training
Through the pretrain run
Model family
Prism.
One job, done exceptionally.
Our specialist line — 13 models, each distilled to do exactly one narrow task well enough to replace a frontier-model call: titling, safety classification, creative writing, roleplay. Nothing general-purpose about them.
- Models
- 13
- Size range
- 135M–8B
- Disciplines
- 4
Model families
Seven families, one philosophy: small, cheap, and good at exactly one thing.
Five are distilled by LoRA from frontier teacher models, and Cybermarine is tuned on our own CyberData set. Quartz is the odd one out — trained from scratch, more on that below.
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.
Amethyst
General-purpose chat — Broader conversational models, including tool-calling variants, built for everyday assistant use rather than one narrow task.
Copal
Agentic tool use — Models trained specifically to decide when and how to call a tool — not just chat, but act.
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.
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.
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.
Aquamarine
Code specialist — A coding-distillation model in active training: generation, debugging, and refactoring across Python, JavaScript, TypeScript, Go, and SQL.
Help shape the next Prism model, or found a bug?
Suggest a model, dataset, or use case, or report an issue — both take under a minute.
Released
Quartz Micro Preview
A mixture-of-experts model, trained from zero
Not a fine-tune — randomly-initialized weights, trained on a single 6GB consumer GPU, then instruction-tuned with a full-parameter SFT pass. Both checkpoints are published on Hugging Face, evaluated honestly either way.
Read the full architecture and progress reportTotal parameters
1,031.0M
~1.03B
Active parameters / token
394.0M
~394M
Hidden size
1,024
What we've learned
A few honest findings from this cycle.
Synthetic data teaches form, not capability
Narrow synthetic distillation reliably fixes formatting and behavior — a title-length spec, suppressed reasoning leakage, correct tool timing — but doesn't reliably add knowledge the base model lacked. Real, human-annotated data is what moved raw accuracy: Prism Safety's BeaverTails recall went from 13% to 79.5% only after mixing in real labeled data.
Base model choice can matter more than fine-tuning
An untuned Qwen3-1.7B outperformed multiple fine-tuned Gemma-based safety classifiers on held-out benchmarks. Picking the right foundation is sometimes a bigger lever than the training run itself.
Held-out evals are non-negotiable
Early evaluation passes that sampled from the same template bank used for training scored close to 100% — because the model had simply memorized the templates. Every eval we publish uses hand-authored or independently-sourced held-out data with verified zero overlap against training.
Small models plus consumer hardware go further than expected
Every fine-tune we've shipped — and the from-scratch 1B-parameter MoE pretrain — runs on Apple Silicon or a single 6GB gaming GPU. No cluster required.
Follow along
New releases and training updates, where you already spend time.
We post models as they ship, the numbers behind them, and what the training runs are doing.
Every model, every weight, on Hugging Face.
MLX and GGUF builds for every release — run them on Apple Silicon or via llama.cpp, LM Studio, and Ollama.