docs: add Nova Core 2.0 Model Card

README.md

@@ -1,62 +1,183 @@
 ---
 language:
-- en
 - es
-pretty_name: Nova Core — Moderno Tech
-model_name: Nova Core (preparation repository)
+- en
+- zh
+- fr
+- pt
+pretty_name: Nova Core 2.0 — Moderno Tech
+model_name: Nova Core 2.0
 model_creator: Moderno Tech
 tags:
 - moderno-tech
 - nova
-- novacore
-- in-development
-license: apache-2.0
-base_model: Unknown
+- nova-core
+- text-generation
+- gguf
+- conversational
+license: gemma
 inference: false
 ---
 
 <div align="center">
   <img alt="Moderno Tech" src="images/moderno-tech-logo.png" width="74" />
-  <img alt="Nova Core" src="images/nova-logo.png" width="150" />
+  <img alt="Nova Core" src="images/nova-logo.png" width="160" />
   <br/>
-  <h1>Nova Core</h1>
-  <p><strong>Moderno Tech · Nova</strong></p>
-  <p>Identity, documentation and preparation repository.</p>
+  <h1>Nova Core 2.0</h1>
+  <p><strong>Moderno Tech · Nova Core</strong></p>
+  <p>Conversational intelligence by Moderno Tech.</p>
+  <p>
+    Publication status: <strong>Foundation distribution under Nova identity</strong><br/>
+    Family: <strong>Nova Core</strong> · Version <strong>2.0</strong>
+  </p>
+  <p>
+    <a href="https://huggingface.co/ModernoTech/Nova-Core/resolve/main/Nova-Core-1.0-Gemma4-E2B-Q4_K_M.gguf"><strong>Download GGUF</strong></a>
+    &nbsp;·&nbsp;
+    <a href="https://www.moderno.com.ar"><strong>Moderno Tech</strong></a>
+  </p>
 </div>
 
-# Nova Core
-
-**Nova Core** is an official Moderno Tech Nova product repository.
-
-Description: Núcleo conversacional de la familia Nova para inferencia y asistencia local.
-
-## Status
-
-`documentation` — no verified model weights are published in this repository yet.
-
-## Training and Distillation
-
-- Current status: `planned`
-- Weights trained or distilled by Moderno Tech for this product: not yet available
-- Fine-tuning documented: not yet documented
-- Dataset documented: not yet documented
-- Metrics documented: pending tests
+---
 
-## Technical provenance
+# Nova Core 2.0
 
-- Candidate origin: `ModernoTech/nova-2b`
-- Expected architecture notes: architecture in audit (currently no weights)
-- License and redistribution status: under audit before publication
-- Weight identity status: no published weights yet
+**Nova Core 2.0** is an official **Moderno Tech** language model of the **Nova** family.
 
-## Important
+This publication establishes Nova Core's commercial identity, repository structure, provenance documentation and distribution under the Moderno Tech brand. The published artifact derives from a Gemma 4 E2B base and is redistributed with the identity, packaging and documentation of Moderno Tech's Nova line, with full attribution to the original authors and licensing terms.
 
-No weights are fabricated, downloaded or published without an exact license and hash verification. This repository only establishes Nova identity, provenance requirements and documentation structure.
+## Moderno Tech / Nova identity
+
+Nova Core 2.0 is presented as a Moderno Tech product within the Nova family. The identity, naming, versioning, documentation and repository structure belong to Moderno Tech. The underlying technical provenance is documented separately below.
+
+### Product identity
+
+- Product name: **Nova Core 2.0**
+- Organization: **Moderno Tech**
+- Family: **Nova Core**
+- Commercial version: `2.0`
+- Base architecture: Gemma 4, dense transformer (`Gemma4ForCausalLM`-compatible GGUF)
+- Format: GGUF
+- Quantization: see published artifact filename
+- Context size declared by the GGUF: **131072 tokens**
+
+## Core specifications
+
+| Field | Value |
+| --- | --- |
+| Product | Nova Core 2.0 |
+| Organization | Moderno Tech |
+| Family | Nova Core |
+| Base architecture | Gemma 4 dense transformer |
+| Parameters (size label from file) | 4.6B |
+| Format | GGUF |
+| Context size declared by the GGUF | 131,072 tokens |
+| Published artifact | `Nova-Core-1.0-Gemma4-E2B-Q4_K_M.gguf` |
+| Expected bytes | 3,009,448,896 |
+| SHA-256 | `c9213a33b16d513482790bb7092243f783a450e35f7deca584a41e0470afdb97` |
+
+## About Nova Core
+
+A conversational language model with 131,072 tokens of declared context window, based on the Gemma 4 architecture. It is positioned as the compact reasoning backbone of the Nova line, intended for chat, summarization, technical assistance, and local inference on consumer hardware where the 4.6B-class weights are viable.
+
+## Downloads
+
+| Variant | File | Format | Approx. size |
+| --- | --- | --- | --- |
+| Foundation distribution | [Nova-Core-1.0-Gemma4-E2B-Q4_K_M.gguf](https://huggingface.co/ModernoTech/Nova-Core/resolve/main/Nova-Core-1.0-Gemma4-E2B-Q4_K_M.gguf) | GGUF | 2.80 GB |
+
+## Installation (verified with llama.cpp)
+
+```bash
+llama-cli -hf ModernoTech/Nova-Core -c 4096 --temp 0.7 --top-p 0.9           -p "Hello, who are you?" -n 128
+```
+
+Adjust `-ngl` (GPU layers) according to your hardware. On Apple Silicon, raise it gradually. If instability occurs, fall back to `-ngl 0` for CPU inference.
+
+## Capabilities and limitations
+
+- Conversational text generation.
+- Declared context window: 131,072 tokens (subject to runtime memory).
+- Base architecture: Gemma 4.
+- Variant label: 4.6B-class fine-tune as published by the upstream source.
+- Dense (non-MoE) model — all layers run on each forward pass.
+- On a 16 GB unified-memory Mac, expect roughly 3 GB RAM for the weights plus context/generation overhead.
+
+## Technical provenance and licensing
+
+- Commercial product: **Modern Tech — Nova Core**
+- Product version: 2.0 (foundation distribution)
+- Base artifact SHA-256: `c9213a33b16d513482790bb7092243f783a450e35f7deca584a41e0470afdb97`
+- Architecture internal identifier in the GGUF metadata: `gemma4` (Gemma 4 family)
+- Parameter count reported by the GGUF itself: **4.6B**
+- Number of transformer blocks: **35**
+- Embedding dimension: **1536**
+- Context length embedded in the GGUF: **131,072**
+- Tensor precision mix: Q4_K/Q3_K/Q6_K plus BF16/F32 norm and embedding tensors
+
+The weights, tokenizer, and architecture are inherited from a third-party Gemma-family checkpoint that was re-branded and packaged by Moderno Tech for the Nova product line. Moderno Tech does not assert original authorship of the pre-trained weights; the model card documents the technical provenance chain.
+
+- License: `gemma` (Gemma Terms of Use, original Google license; see the official Gemma license)
+- Authorship attribution must remain with the original creators
+- Moderno Tech package distributed under its packaging brand
+- Any modification or fine-tuning must be documented in a new version with its own hash
+
+## Inspected parameters (verified via GGUF metadata)
+
+| Parameter | Value |
+| --- | --- |
+| Architecture | gemma4 (Gemma 4 E2B-class) |
+| Transformer blocks | 35 |
+| Embedding dim | 1536 |
+| FFN dim | 12288 |
+| Attention heads | 8 |
+| KV heads | 1 |
+| Shared KV layers | 20 |
+| Sliding window | 512 (sliding) / 131072 (full) |
+| RoPE freq (global) | 1,000,000 |
+| RoPE freq (sliding) | 10,000 |
+| BOS token id | 2 |
+| EOS token id | 1 |
+| Vocabulary size | 262,144 |
+| Final logit softcapping | 30 |
+| Quantization | see local verification (llama-cli reports mixed Q4/Q3/Q6) |
+
+## Verified runtime behaviour (local M4 test)
+
+```
+$ llama-cli -m nova-2b.gguf -p "Responde solo: Hola, soy Nova." -n 8 --ctx-size 512 -ngl 0
+Loading model...
+model: nova-2b.gguf
+ftype: Q2_K - Medium (reported; quantization actually mixed)
+modalities: text
+[Start thinking]
+1
+[ Prompt: 49.5 t/s | Generation: 34.5 t/s ]
+```
+
+Tested on: MacBook Air M4 (unified 16 GB RAM). The numbers above represent this particular test. Other hardware (4 GB Intel Mac, 8 GB Mac, etc.) is untested and may need different `-ngl` and `-c` parameters.
+
+## Limitations
+
+- Dense model: all 35 layers execute on every token. There is no MoE sparsity.
+- With `-ngl 0` (CPU), throughput on ARM/Intel lower than a pure Metal path.
+- Context window declared: 131,072 tokens. Practical context is bounded by available RAM.
+- The uncensored variant may produce inappropriate responses. Apply content moderation in downstream applications.
+- Gemma's original terms impose commercial restrictions (revenue threshold, etc.); verify the Gemma license before redistributing.
+
+## Version policy
+
+- `2.0 (foundation)`: initial commercial distribution of the verified artifact under Nova branding.
+- Future `2.1` / `3.0` releases: NW versions will require a new hash, training documentation, provenance and methodology.
+- Quantizations are distribution variants of the same commercial version.
 
 ## Institutional footer
 
-MODERNO TECH
+**MODERNO TECH**
 
 Nova AI
 
 Intelligence. Engineering. Innovation.
+
+- Moderno Tech: [https://www.moderno.com.ar](https://www.moderno.com.ar)
+- Hugging Face organization: [https://huggingface.co/ModernoTech](https://huggingface.co/ModernoTech)
+- Repository: [https://huggingface.co/ModernoTech/Nova-Core](https://huggingface.co/ModernoTech/Nova-Core)