Vertex 0.6 100M — 8192-ctx Instruct
Instruction-tuned chat model of the Vertex 0.6 family by Vertex Research. Built on the 8192-context base (10.01B pretrain tokens + three-stage rope ladder), then SFT'd on ~44.5K conversations: general instruct data, QA/tutoring, function/tool calling, and self-identity.
Model details
| Parameters | 96.75M (tied embeddings) |
| Architecture | Qwen3-based transformer |
| Context length | 8192 (RoPE theta 1M) |
| Chat format | ChatML (`< |
| Tool calling | <tool_call> JSON blocks, system-prompt function definitions |
| EOS | </s> (2) and `< |
Usage
from transformers import AutoTokenizer, AutoModelForCausalLM
repo = "VertexResearch/Vertex-0.6-100M-8192-Instruct"
tok = AutoTokenizer.from_pretrained(repo)
model = AutoModelForCausalLM.from_pretrained(repo)
msgs = [{"role": "user", "content": "Who are you?"}]
text = tok.apply_chat_template(msgs, add_generation_prompt=True, tokenize=False)
ids = tok(text, return_tensors="pt", add_special_tokens=False).input_ids
out = model.generate(ids, max_new_tokens=100, eos_token_id=[2, 6])
print(tok.decode(out[0, ids.shape[1]:], skip_special_tokens=True))
For tool calling, put function definitions in the system prompt; the model emits
<tool_call>{"name": ..., "arguments": ...}</tool_call> and consumes results in
<tool_response> blocks.
Training
SFT with TRL on top of Vertex-0.6-100M-8192-ctx-Base: 2 epochs, lr 3e-4 cosine, bf16, max length 2048. Final eval loss 1.466.
Limitations
These models are not the most coherent yet and need more tuning: expect rambling, repetition, and inconsistent answers, especially over longer generations.
97M parameters: fluent chat and working tool-call syntax, but frequent factual errors, shallow reasoning, and rambling explanations. Not for production use. Knowledge cutoff ~April 2024.