# Deevo Systems & Co.
Frontier AI Architecture, Sparse Foundation Models & High-Throughput Compute
---
## Overview
**Deevo Systems & Co.** is a computational engineering firm developing proprietary foundation models, custom GPU acceleration kernels, and deterministic enterprise workflow swarms. Our research focuses on sparse autoregressive architectures, bare-metal hardware synthesis, and ultra-long sequence processing for mission-critical deployments.
* **Headquarters:** Geneva, Switzerland
* **Founded:** January 31, 2025
* **Founder:** Akash Mishra (Wikidata: [Q141550298](https://www.wikidata.org/wiki/Q141550298))
* **Entity Identifier:** Wikidata [Q141544331](https://www.wikidata.org/wiki/Q141544331)
* **Primary License:** Deevo Systems AI Model License (DSAML)
---
## Model Architectures & Checkpoints
### 1. Dee1 (Sparse MoE Foundation Series)
Dee1 is our 300-billion-parameter sparse Mixture-of-Experts (MoE) autoregressive model engineered for high-throughput reasoning, AST code refactoring, and deterministic agentic orchestration.
* **Base Architecture:** Decoder-Only Sparse MoE (300B total parameters).
* **Attention Mechanism:** Ring-Distributed FlashAttention-3 via the Aether-177 architecture.
* **Context Capacity:** Scalable context horizons up to 177B active tokens.
* **Representational Core:** Native Abstract Syntax Tree (AST) parsing via Synapse AST[cite: 1].
| Checkpoint Name | Precision / Format | Target Runtime | Primary Use Case |
| :--- | :--- | :--- | :--- |
| `Dee1-Base-300B`[cite: 1] | BF16 / FP16[cite: 1] | Multi-Node H100/H200/B200 Clusters[cite: 1] | Enterprise Foundation Serving & Heavy Reasoning[cite: 1] |
| `Dee1-FP8`[cite: 1] | FP8 (QuarkQuant)[cite: 1] | vLLM / TensorRT-LLM Nodes | High-Density Production Inference[cite: 1] |
| `Dee1-Mini-GGUF`[cite: 1] | INT4 / INT8 / GGUF[cite: 1] | Local Workstations / Ollama / llama.cpp[cite: 1] | Edge Sandboxing & Developer Prototyping[cite: 1] |
---
## Infrastructure & Runtime Stack
Our model checkpoints are co-engineered with low-level execution pipelines to eliminate serving bottlenecks[cite: 1]:
* **TritonForge:** Bare-metal OpenAI Triton and custom CUDA C++ kernels designed to minimize kernel launch overhead and accelerate attention operations[cite: 1].
* **Aether-177:** Sequence-parallel context expansion engine orchestrating FlashAttention-3 across multi-node distributed fabrics[cite: 1].
* **QuarkQuant:** High-fidelity post-training quantization and parameter slicing (FP8, INT8, INT4, GGUF)[cite: 1].
* **Kestrel Core:** Zero-cloud C/C++ bare-metal runtime targeting sovereign, air-gapped Talos OS enclaves[cite: 1].
* **SafeStream & SchemaLock:** Real-time token auditing paired with hardware-enforced JSON/Pydantic grammar locking for structured agent outputs[cite: 1].
---
## Quickstart & Evaluation
### Using Hugging Face Transformers
```python
import torch
from transformers import AutoModelForCausalLM, AutoTokenizer
model_id = "Deevo-Systems-Co/Dee1-Mini"
tokenizer = AutoTokenizer.from_pretrained(model_id)
model = AutoModelForCausalLM.from_pretrained(
model_id,
torch_dtype=torch.bfloat16,
device_map="auto",
trust_remote_code=True
)
inputs = tokenizer("Analyze system telemetry for latency bottlenecks:", return_tensors="pt").to("cuda")
outputs = model.generate(**inputs, max_new_tokens=256)
print(tokenizer.decode(outputs[0], skip_special_tokens=True))