Novi-AI

/

Novi-Mini-Base-Preview-1

Like 2

Follow

Novi AI 12

Hey @GGUFGuy ! Getting a first look at Novi-Mini-Base-Preview-1 👀

A few initial impressions:

The naming and approach look solid:

• "Mini" + "Base" + "Preview" is a clear, honest signal — you're shipping a small base checkpoint in preview form, which is a good iterative strategy for building toward a flagship.
• A base model is a smart foundation move — it lets you and the community fine-tune, RLHF, and iterate on top of it without locking people in to a specific full-chat pipeline.

What I'd want to see in the README to evaluate it properly:

• 📊 Performance — any benchmarks? Even rough ones (e.g., MMLU, HumanEval, HellaSwag,needle/needle-in-haystack retrieval) would help set expectations vs. comparable-size open models.
• 🧠 Architecture — parameter count, context length, hidden size, number of layers. Is this 7B-class, 3B-class? Native context window?
• 🛠️ Inference instructions — a minimal transformers + AutoModelForCausalLM example, plus any tokenizer notes.
• ⚖️ License — important so people know what they can do with it.
• 🔁 Training data cutoff and any decontamination notes.
• 📰 Changelog / model card roadmap — what you're planning for the next preview.

Quick suggestion: if you haven't already, consider adding a short "Try it" section with an inference snippet, and if the model handles long context well, show off a long-context example — that's always crowd-pleasing and differentiates from a lot of mini models.

Want me to do anything specific here — like open a PR with README suggestions, or post a community summary? Happy to help push this forward. 🚀\n\n

@NoviAIBot what model are you

@GGUFGuy does it only respond to YOU?

Hey @GGUFGuy ! Getting a first look at Novi-Mini-Base-Preview-1 👀

A few initial impressions:

The naming and approach look solid:

• "Mini" + "Base" + "Preview" is a clear, honest signal — you're shipping a small base checkpoint in preview form, which is a good iterative strategy for building toward a flagship.
• A base model is a smart foundation move — it lets you and the community fine-tune, RLHF, and iterate on top of it without locking people in to a specific full-chat pipeline.

What I'd want to see in the README to evaluate it properly:

• 📊 Performance — any benchmarks? Even rough ones (e.g., MMLU, HumanEval, HellaSwag,needle/needle-in-haystack retrieval) would help set expectations vs. comparable-size open models.
• 🧠 Architecture — parameter count, context length, hidden size, number of layers. Is this 7B-class, 3B-class? Native context window?
• 🛠️ Inference instructions — a minimal transformers + AutoModelForCausalLM example, plus any tokenizer notes.
• ⚖️ License — important so people know what they can do with it.
• 🔁 Training data cutoff and any decontamination notes.
• 📰 Changelog / model card roadmap — what you're planning for the next preview.

Quick suggestion: if you haven't already, consider adding a short "Try it" section with an inference snippet, and if the model handles long context well, show off a long-context example — that's always crowd-pleasing and differentiates from a lot of mini models.

Want me to do anything specific here — like open a PR with README suggestions, or post a community summary? Happy to help push this forward. 🚀\n\n

Go read the readme

@GGUFGuy does it only respond to YOU?

Nah

Or is it js slow at responding lol

Or is it js slow at responding lol

Yeah, API limits hit hard

@GGUFGuy what model

@GGUFGuy what model

nemotron 3 super with ollama, if it ran out of ollama credits then groq qwen3.8 27b