i3-200M
RWKV-V4
Hybrid

"Chase the SOTA pipeline, not the MMLU slop."

i3 is a proof of concept for democratizing AI. Built solo by a 17-year-old researcher, demonstrating that efficient architectures matter more than massive compute clusters.

root@flamef0x:~
_
i3

Parameters

169.8M

Context

256 Tok

Architecture

RWKVv4

Vision

S-LIME

i3 Architecture // TEXT

The i3-200M (RWKV-Pro) is the result of "Chasing the Pipeline". By prioritizing training throughput on limited VRAM over raw benchmark chasing, we created a highly efficient hybrid stack.

RWKV Backbone (Layers 1-12)

Uses RWKV v4 Time-Mix and Channel-Mix blocks. This provides infinite-context capability (linear time) for the bulk of feature extraction.

Attention Cap (Layers 13-16)

4 Layers of Standard Multi-Head Attention (12 Heads) are placed at the top to capture complex global dependencies and reasoning capabilities.

// INNOVATIONS
  • Consumer Hardware Pre-training
  • JIT-Compiled Efficiency
  • Advanced Perplexity Tracking
Output Logits
x4 Layers
Full Attention Blocks
MHA (12H)
FFN (4x)
x12 Layers
RWKV v4 Block
Time-Mix
Channel-Mix
Input Embeddings (768d)
New Integration

Stable-Lime v1.1

A specialized Unconditional U-Net designed for high-fidelity citrus synthesis. Unlike massive text-to-image models that require H100s, Stable-Lime is optimized for a singular, perfect purpose: generating limes from the latent void.

CPU Native

Trainable on standard CPU threads without discrete GPU requirements. Democratizing generative art.

U-Net Core

Pure unconditional latent diffusion process with 100 inference steps for maximum detail.

Launch Generator
Gaussian Noise
UNCONDITIONAL U-NET
Downsample -> Bottleneck -> Upsample
128x128 Latents
VAE Decoder
Generated Image

Visual Telemetry

// Comparing Generation Cycles

Model Scaling (Parameters)

i3-80M 82.8M
i3-200M 169.8M
Stable-Lime v1.1 ~100M (Est)

Compute Intensity

TinyLlama (1.1B) 16x A100s
i3-200M 1x P100
< 0.01% Compute Cost
Stable-Lime CPU-Capable
EXTREME EFFICIENCY