Most AI companies in India wrap external frontier APIs. We took the harder path: building and training our own sovereign model from pure scratch.
Meet LuxurAI Nova-Mini V1 — an ultra-compact 57,192,128 parameter transformer engineered to execute deterministic code and sysadmin tasks directly on edge hardware.
Why 57 Million Parameters?
Frontier models like Claude 3.5 or GPT-4 require immense cloud clusters and hundreds of watts of power. But for specialized automation tasks — such as desktop terminal automation, LeetCode verification, and arithmetic execution — massive weight sizes are wasteful.
At 57.2M parameters (~109.1 MB FP16 weights on disk), Nova-Mini V1 can run completely locally on low-end laptops, mobile devices, or edge micro-servers with sub-100ms per-token latency when paired with our KV-cache.
Model Architecture & Technical Specifications
- Parameter Count: 57,192,128 parameters
- Transformer Layers: 22 attention layers
- Hidden Dimension: 448
- Multi-Query Attention (MQA): 7 Query Heads, 1 Key-Value Head (7Q/1KV)
- Activation Function: SwiGLU with tied embeddings
- Vocabulary Size: 16,384 BPE tokens
- Convergence Metrics: SFT V2.1 Loss 0.4421 | Perplexity 2.65 | Validation Accuracy 82.10%
- Hardware: Trained on dual Tesla T4 GPUs
The Architecture Secret: Program-of-Thought (PoT) Tool Delegation
In our live benchmark suite, raw parametric mental arithmetic in small models frequently hallucinates. Instead of attempting 6-digit multiplication in raw weights, Nova-Mini emits executable Python REPL sandboxed code:
# Prompt: Multiply 1000453 x 333
<thought>
The user requires multiplication of two large integers.
Parametric calculation has variance; delegating to deterministic REPL.
</thought>
```python
result = 1000453 * 333
print(result)
```
# Sandbox Output: 333150849What's Next?
We are actively training Phase 2 (22K long-context annealing) and exporting GGUF Q4_K_M weights for llama.cpp / Ollama compatibility. Track the active development at luxurai.in.