Investors

Small models, ordinary hardware.

Two ideas for LLMs that run on CPU, not in a GPU data centre. Both are early. Both are grounded in things already running on our own server — and this page says plainly what is proven, what is not, and what each idea cannot do.

Read this first

An honest page, on purpose.

No projections live on this page. No valuation, no round size, no forecast, no market-size chart. If you want numbers, ask us — we will show you the evals, the live sites, and the code.

Proven

What already works

Two 278M-parameter models, fine-tuned and serving live on CPU. Two retrieval sites answering from real sources. A decision-model API already running on CPU. These are running systems, not slides.

Not proven

What does not exist yet

Anything bigger than 278M on CPU. A newsroom model that drafts publishable copy. Any revenue, any market, any timeline. Both ideas below are early, and the second one is still only a written sketch.

Idea 01 — In build

A family of small models that run on CPU.

Task-specific models, small enough to run on an ordinary server CPU — no GPU farm, no per-token API bill, and data that never has to leave our own machine. Telugu verse is the proof: two 278M models already finish a poem or a song from its first line, live, on CPU.

278M
Parameters per model
~285 MB
Model file (GGUF Q8_0)
2
Models live on CPU today
1,024
Token context window
What is proven

Trained, shipped, serving

The poem model trained in 1,539 steps, about two hours on a free Kaggle T4, final loss 2.58. The song model took 126 steps, about eighteen minutes, loss 2.88 down to 2.45. Both export to GGUF, run under llama.cpp in Docker on our server, and answer live at poems.socialnews.xyz and songs.socialnews.xyz. Training at this scale cost no GPU rental at all — the work was data, eval, and serving. Sampling settings (temperature 0.8, top-p 0.9, repetition penalty 1.15) are what make the output usable.

What it cannot do

The ceiling, stated plainly

On 40 held-out poems the fine-tuned model recalled exactly zero verbatim — it learns style and form, not unseen text. That is a real ceiling, not a footnote. Long coherent prose is the wall: a 1,024-token context, stories that repeat after about 150 words, logic that slips. Greedy decoding loops. And CPU inference is slow per token — fine for short outputs, slow for long ones.

What money would do

Bigger small models in the 1B class — untested on our CPU so far, so that is an experiment, not a promise. More languages and domains, a real eval harness, serving hardened beyond a single box, and data licensed properly: film lyrics are copyrighted, so our song data stays private and research-only.

llama.cppGGUFDockerKaggle free tierHugging Face
Idea 02 — Sketch only

A newsroom model that drafts, never publishes.

Three small servers: feeds for what is new, knowledge for semantic search over the archive with a duplicate check, and drafts for where the copy goes. Draft-only by design — no publish tool, a human reviews every draft. Written as a sketch in November 2026; not built, and parked until we choose to build it.

~75%
Tool-call accuracy, 7B models
~68%
Tool-call accuracy, 1.5B models
~40%
4 tool steps at 80% each
0
Publish tools, by design
What is proven

The retrieval half already runs

The grounded-answers half of this idea is live today: Sanatana Dharma (448 chunks) and Vemana (572 poems) answer only from retrieved passages, citing the exact page or poem number. The sketch itself is written — three servers, draft-only, no publish tool — in our small-mcp-core prototype. The drafting half has not been built.

What it cannot do

The bottleneck is reliability

Small models are unreliable tool users: about 75% per tool call at 7B, about 68% at 1.5B, and a cliff below that. Four tool steps at 80% each is about 40% end-to-end — that is the bottleneck, and it is why the design is draft-only. Without retrieval grounding a CPU model drafting news would hallucinate, and it cannot replace reporters or verify facts on its own.

What money would do

Build the knowledge server first — the smallest test, search plus duplicate check over the archive — then an eval harness, then feeds, then drafts, in that order. Each step is gated on the one before it working.

Retrieval groundingMCPDraft-onlyHuman review
Risks

What could make this fail.

Five risks, in plain language. None of them is hidden in a footnote.

Risk 01

Frontier labs move fast

Open-weight models improve every few months. A 278M model is not a moat — the advantage, if any, is in the data, the evals, and the serving, not the weights.

Risk 02

Small models hallucinate

Grounding and human review are load-bearing, not optional. Any use without retrieval or review will produce confident wrong answers.

Risk 03

CPU is a ceiling

Bigger models get slower on CPU, not just better. There is a size where the idea stops working, and we have not found it yet.

Risk 04

Data licensing

Song lyrics are copyrighted; news text belongs to publishers. Some data can only ever be private and research-only, which limits what can be shipped.

Risk 05

Distribution is the hard part

Building the model is the easy half. Getting anyone to use it — in a newsroom, on a device, in a language the big labs ignore — is the work that decides whether this matters.

Talk to us

Bring your own diligence.

We will show you the evals, the live sites, and the code — including the evals that came back zero. That is the point of this page.

See the work behind this page: Telugu AI · poem model · song model · news-model sketch