christud/superchat-35b-a3b

Superchat 35B-A3B

Sovereign AI. On your machine. Zero cloud.

Overview

Superchat is a 35B parameter AI model (3B active per token via MoE) with:

• Tool calling — Read/write files, run commands, edit code
• 1M token context — Natively, extensible to 10M+ via disk retrieval
• 201 languages — Including 11 Indian languages
• Claude-level quality — Distilled from Claude Opus 4.6 via Chimere LoRA
• Runs on laptops — Only 3B params active, fits in 16GB RAM

Architecture

Property	Value
Total Parameters	35 billion
Active per Token	3 billion (MoE sparse)
Context Window	1,000,000 tokens
Languages	201
Base Model	Qwen3.5-35B-A3B

LoRA Stack

1. Chimere — Distilled from Claude Opus 4.6 (tool-calling, agentic reasoning)
2. CLI Agent — Terminal workflows, ML operations
3. Superchat Identity — Indian languages, custom knowledge, branding
4. Finance — Financial chain-of-thought reasoning
5. Vision/OCR — Document understanding
6. Deep Thinking — Enhanced reasoning via Fragmented Training

Indian Languages

Hindi, Tamil, Telugu, Malayalam, Kannada, Bengali, Marathi, Gujarati, Punjabi, Odia, Assamese, Urdu, Sanskrit, Nepali

Usage

from transformers import AutoModelForCausalLM, AutoTokenizer

model = AutoModelForCausalLM.from_pretrained("christud/superchat-35b-a3b")
tokenizer = AutoTokenizer.from_pretrained("christud/superchat-35b-a3b")

messages = [
    {"role": "system", "content": "You are Superchat, a sovereign AI by Christudas Philipose."},
    {"role": "user", "content": "Hello! What can you do?"},
]

Creator

Christudas Philipose superchat.in Made in India, for the world.

License

Apache 2.0