Words become pieces.
The tokenizer splits a sentence into small pieces and assigns each piece a number. Those numbers are the model's input.
An independent language-model project · 01 / 04
A language model you can look inside.
OmegaMind is a 152.8-million-parameter language model with its own training pipeline and a custom C++17 inference engine. It generates text on a CPU, using system memory for its model weights, working data and attention cache.
CPU INFERENCE · AVX2 · FMA3 · OPENMP · SYSTEM RAM
Inside the machineInference path / 01
01 / Storagemodel.binTrained weights on disk
02 / Host memorySystem RAMMapped weights · KV cache · buffers
03 / ComputeCPUC++17 · AVX2/FMA · OpenMP
Memory mappingLoad & reuseCalculate
NO GPU INFERENCEONE TOKEN AT A TIME
152.8MModel parameters · approx.
16Transformer layers
12Attention heads
FP32Exported weight format
Start with the idea
It takes a question apart, then builds an answer.
A language model does not read a sentence in the same way a person does. It turns text into numbers, processes those numbers through learned layers, and predicts what should come next.
The CPU works through the numbers.
OmegaMind looks up learned values, passes them through 16 Transformer layers and reuses information held in system RAM.
A response appears one piece at a time.
The model selects a next piece, converts it back into text and repeats the process until the answer is complete.
The system behind the answer
Three parts. One continuous process.
The design is intentionally CPU-first. The model file, its working memory and the calculations each have a distinct place in the system.
Weights on storage
The trained parameters live in model.bin. C++ maps the file into its address space instead of copying the entire file into another weight buffer.
→mapped into memory
MODEL.BINMMAP
System memory
Host RAM holds the session's K and V attention cache and temporary buffers used while the CPU performs each calculation.
K / V + working buffers
HOST RAM96 MiB K/V ALLOCATION PER SESSION*
CPU execution
AVX2 and FMA process groups of floating-point values. OpenMP distributes suitable loops across CPU threads.
parallel work
C++17AVX2 · FMA · OPENMP
* Based on 16 layers × 1,024 positions × 768 dimensions × FP32. This is the allocated K/V buffer capacity, not a claim about physical RAM residency.
A closer look · Interactive walkthrough
Follow one question through OmegaMind.
This is the original token-journey experience, kept as its own self-contained section. It is loaded inline so it can run inside the Hugging Face Space frame. Scroll inside it, pause the walkthrough or move through the stages at your own pace.
OMEGAMIND / ONE TOKEN JOURNEYSCROLL WITHIN THE FRAME TO EXPLORE
The original interactive walkthrough is displayed directly in this frame.Open the walkthrough in a new tab
Your turn
Ask the model. Watch it answer.
Enter a question and send it to the OmegaMind inference engine running behind this page. The reply is streamed back as the model generates it.
Inference engineC++17 / CPU
Input tokenizerGPT-2 TIKTOKEN
Parallel executionOPENMP
MemoryHOST SYSTEM RAM
Talk to OmegaMindChecking engine
A question starts the process.Ask something, then watch the answer arrive from the model running on the Hugging Face Space.
Waiting for a question
Open for opportunities
Interested in building systems that go beyond the demo?
I'm open to conversations around inference engineering, efficient model execution, systems programming and applied AI. If you are building in this space, I'd be glad to connect.
Visit NanoMindAnother systems project↗
Source on GitHubOmegaMind repository↗
Connect on LinkedInProfessional profile↗
Send an emailDirect contact↗
For collaborations, roles or technical conversationssumeetkhatri1771@gmail.com