OPEN-WEIGHT AI · BUILT IN INDIA

Most people
fine-tune models.
We build them.

Tantra AI Labs builds open-weight language models from scratch — every layer, every weight, every token — plus the reasoning datasets and production infrastructure to run them.

7B
parameters
193K
reasoning triples
RTX 4090
PyTorch Grouped-Query Attention RoPE SwiGLU LoRA / QLoRA Unsloth BF16 Mixed Precision Custom BPE Tokenizer Chain-of-Thought vLLM Chain-of-Thought 2× RTX 4090 PyTorch Grouped-Query Attention RoPE SwiGLU LoRA / QLoRA Unsloth BF16 Mixed Precision Custom BPE Tokenizer Chain-of-Thought vLLM Chain-of-Thought 2× RTX 4090
01 — MODELS

Open-weight, from scratch.

Base · Reasoning

AstraGPT-7B

FLAGSHIP

7B decoder-only LLM. Custom BPE tokenizer, GQA (28Q/4KV), RoPE, SwiGLU. Native <think> chain-of-thought. 32K context.

7B
Params
32K
Context
Apache-2.0
License
View on Hugging Face →
Code · Reasoning

AstraGPTCoder-7B

Coding-specialized sibling. LoRA fine-tuned for code generation with think-tag reasoning. Runs in ~6GB VRAM at 4-bit.

AstraGPT-7B
Base
~6GB
VRAM (4-bit)
Apache-2.0
License
View on Hugging Face →
ARCHITECTURE

Every layer, hand-built.

No AutoModel. No copy-paste. AstraGPT's transformer, BPE tokenizer, and supervised fine-tuning loop were all written from first principles in PyTorch.

Grouped-Query Attention28 query / 4 KV heads — 7× less KV memory
RoPEθ = 1,000,000 — long-context positional encoding
SwiGLU FFNintermediate 18,944 — beats GELU/ReLU
RMSNormpre-norm, ~30% faster than LayerNorm
Token IDs
  │
  ▼  Token Embedding [64k → 4096]
  ▼  ×28 AstraGPT Blocks
┌─────────────────────────┐
│ RMSNorm (pre-norm)      │
│ → GQA  (28Q / 4KV)      │
│   RoPE θ=1M · KV-cache  │
│ → residual              │
│ RMSNorm                 │
│ → SwiGLU FFN            │
│ → residual              │
└─────────────────────────┘
  ▼  Final RMSNorm
  ▼  LM Head [4096 → 64k]
  ▼  Logits → next token
UNDER THE HOOD

The anatomy of AstraGPT-7B.

28 transformer blocks, grouped-query attention with 4× fewer KV heads, rotary embeddings at θ=1M and a SwiGLU feed-forward — every component implemented and understood from first principles.

28
layers
28Q/4KV
GQA heads
32K
context
astragpt/architecture
Token IDs
  │
  ▼  Token Embedding [64k → 4096]
  ▼  ×28 AstraGPT Blocks
┌─────────────────────────┐
│ RMSNorm (pre-norm)      │
│ → GQA  (28Q / 4KV)      │
│   RoPE θ=1M · KV-cache  │
│ → residual              │
│ RMSNorm                 │
│ → SwiGLU FFN            │
│ → residual              │
└─────────────────────────┘
  ▼  Final RMSNorm
  ▼  LM Head [4096 → 64k]
  ▼  Logits → next token
02 — DATASETS

Reasoning, at scale.

0triples

AstraGPT Reasoning Dataset

Instruction–reasoning–answer triples with explicit cognitive traces. Apache-2.0.

Download dataset →
0Q&A pairs

High-Reasoning Dataset v1

Multi-AI knowledge-distilled. Deep CoT, math proofs, production code, edge cases. MIT.

Download dataset →
03 — INFRASTRUCTURE

From training rigs to production gateways.

01

AstraGPT Training

Bare-metal VPS, 2× RTX 4090 (48GB). Custom PyTorch loop — BF16, grad accumulation, cosine LR, W&B logging.

02

Astra AI Gateway

Production LLM gateway. 40+ models, multi-channel routing, fallbacks, HTTP/2 zero-buffer streaming, cost analytics.

03

vLLM Serving

Self-hosted inference with prefix caching. Optimized connection pooling cuts per-request latency ~40%.

Try the gateway — one OpenAI-compatible endpoint, 40+ models

$curl https://api.aisubscription.shop/v1/chat/completions
04 — THE LAB

Built by one engineer.

Tantra AI Labs is founded by Aditya Wakharkar — an AI infrastructure and cybersecurity engineer from Nanded, India. From training 7B models on a private GPU rig to shipping production gateways and ethical-hacking tools, the lab operates on one belief: open weights, built honestly.