VertexAGI/amethyst-1.5-mini

Amethyst 1.5 Mini

Amethyst 1.5 Mini is a general-purpose chat model fine-tuned from Gemma 3 4B IT, expanding on 1 Mini's plain-chat focus with reliable web-search tool-calling: knowing when to call a search tool, when not to, how to decompose multi-part questions into multiple queries, how to refine a query after a bad result, and how to summarize retrieved snippets into a grounded, cited answer.

Part of the Amethyst family of chat-optimized models.

Model Details

Developed by	Independent research project
Base model	google/gemma-3-4b-it
Fine-tuning base checkpoint	mlx-community/gemma-3-4b-it-qat-4bit
Architecture	Gemma 3, 4B parameters (dense, decoder-only transformer)
Fine-tuning method	LoRA (rank 8, scale 20.0, 16 layers)
Fine-tuning framework	MLX / mlx-lm, on Apple Silicon
Language	English
License	Gemma Terms of Use

Training Data

Amethyst 1.5 Mini was fine-tuned on 4,987 examples (4,589 train / 398 validation) synthetically generated via knowledge distillation from a mix of nvidia/nemotron-3-super-120b-a12b and nvidia/nemotron-3-ultra-550b-a55b through the NVIDIA NIM API — 5x the training data of 1 Mini's original 1,122 examples.

Category breakdown:

Category	Count	Purpose
General chat	3,264	Explanation, reasoning, code, planning, creative writing, extraction, brainstorming, math, sentiment, translation
Follow-up	438	Multi-turn conversations with elliptical second turns ("what about the other one?")
Summarize	374	Grounded, cited answers synthesized from retrieved search snippets
Single-query search	305	One user question -> one web_search call
Refine-after-bad-result	183	Query -> weak results -> refined follow-up query
Multi-query decomposition	149	One question requiring multiple distinct search queries
No-search (negative class)	274	Answerable from knowledge alone — teaches the model not to search for everything

The no-search category is the key addition over a naive "always search" approach: without negative examples, a tool-calling model tends to reach for search even on trivial questions like "what's 7*8".

Training Procedure

• Method: Supervised fine-tuning via LoRA (rank 8, scale 20.0, dropout 0.0, 16 layers)
• Optimizer: Adam, learning rate 1e-5 (constant schedule)
• Sequence length: 2,048 tokens
• Batch size: 2 (reduced from 4 mid-run after peak memory climbed to 13GB+ on a 16GB machine; the smaller batch keeps training stable and unattended-safe)
• Gradient checkpointing: enabled
• Training steps: 3,250 iterations, with validation every 200 steps
• Final validation loss: 0.652 (down from 0.807 at the point training was resumed with the reduced batch size)

Evaluation

A hand-authored held-out set (26 prompts spanning general chat and explicit tool-call correctness, zero overlap with the training scenario bank) compared base Gemma 3 4B IT against Amethyst 1.5 Mini:

Metric	Base Gemma 3 4B IT	Amethyst 1.5 Mini
Tool-call correctness (16 prompts: 8 should-call, 8 should-not-call)	5/16 (31%)	16/16 (100%)
General chat quality (10 prompts, blind pairwise judge)	2 wins	2 wins, 6 ties

The base model calls web_search inconsistently — sometimes searching for things it already knows ("what's the boiling point of water"), sometimes failing to search for genuinely current information ("is there a new iPhone this year"). Amethyst 1.5 Mini gets this right on every held-out prompt, while general chat quality holds steady rather than regressing.

Tool-calling format

The model was trained to emit tool calls in this format:

<tool_call>
{"name": "web_search", "arguments": {"queries": ["query text"]}}
</tool_call>

A tool result is expected back as a <tool_result> block; the model then produces a grounded, cited final answer.

Intended Use

Amethyst 1.5 Mini is intended as a lightweight, general-purpose conversational assistant with tool-use awareness — for experimentation, research into small-scale distillation pipelines, and hobbyist deployment with a search backend wired up. It is not intended for high-stakes, safety-critical, or production use.

Limitations

• Trained on a moderately sized (~5,000-example) synthetic dataset — behavior can be inconsistent outside the categories represented in training.
• Inherits the general limitations and knowledge cutoff of its base model, Gemma 3 4B IT.
• Distilled from Nemotron teacher models without human review of every example; synthetic-data artifacts may be present.
• Tool-calling was trained against a single tool (web_search) with a fixed schema — behavior with other tool schemas is untested.

Formats available

This repo includes both:

Format	File	Notes
MLX (4-bit)	model.safetensors + config	For Apple Silicon via mlx-lm
GGUF (Q4_K_M)	amethyst_1_5_mini_Q4_K_M.gguf	For llama.cpp and compatible runtimes (LM Studio, Ollama, etc.)

Usage — MLX

from mlx_lm import load, generate

model, tokenizer = load("VertexAGI/amethyst-1.5-mini")

messages = [{"role": "user", "content": "What's the current inflation rate in the US?"}]
text = tokenizer.apply_chat_template(messages, add_generation_prompt=True, tokenize=False)

response = generate(model, tokenizer, prompt=text, max_tokens=500)
print(response)

Usage — GGUF (llama.cpp)

llama-cli -hf VertexAGI/amethyst-1.5-mini -m amethyst_1_5_mini_Q4_K_M.gguf -p "Explain how vaccines train the immune system, in simple terms."

Citation

If you reference this model, please cite it as:

@misc{amethyst15mini,
  title  = {Amethyst 1.5 Mini},
  author = {Independent research project},
  year   = {2026},
  note   = {LoRA fine-tune of Gemma 3 4B IT with web-search tool-calling, distilled from a Nemotron-3-Super/Ultra mix}
}

This model is built on Gemma and subject to the Gemma Terms of Use.