DEV Community: Kata Omel

I built a Can I run this LLM? site where every VRAM number is measured, not guessed

Kata Omel — Sat, 19 Sep 2026 09:21:32 +0000

The problem: five sites, four different answers

When Kimi K3 dropped, I searched "kimi k3 vram requirements" like everyone else.
Five popular pages gave me:

594 GB
1.51 TB
1.68 TB
"you can't run it, period"
one page said Q4 fits in "about 800 GB", another said 1.2 TB

They can't all be right. And they're not all wrong either — they're computing
different things with the same formula: params × bits / 8. That formula
ignores the embedding table (often unquantized even in Q2 builds), attention
tensors kept in higher precision, the output head, and container overhead.
For a 1T-parameter MoE like Kimi K3, those "small" omissions add up to
hundreds of gigabytes.

The fix: stop computing, start measuring

Every GGUF quant of every popular model is already published as a file on
Hugging Face. The file size IS the answer — it's what you download, what loads
into memory. No formula needed.

So I built ModelFit:

fetch.py walks the HF API (public, no key): trending GGUF repos + top by downloads + hype keywords. For each repo it walks the tree (big models store quants in Q4_K_M/ subdirs, shards like -00001-of-00012.gguf get summed), and pulls config.json for exact KV-cache math.
build.py computes need = weights + KV cache + overhead, and renders one page per model: quant-by-quant table, verdicts for every GPU from RTX 3060 to Mac Studio 512GB, an interactive "type your VRAM" checker, FAQ with JSON-LD.
KV cache is exact when config.json is available (GQA formula: 2 × layers × kv_heads × head_dim × ctx × 2 bytes), otherwise a flat +20%/min-1.5GB heuristic — and the page says which one it used.
A daily cron re-fetches, re-tests (formula gate against llama.cpp reference numbers), rebuilds, deploys to Vercel, and pings IndexNow. New model release → page live within 24h, usually before the blog posts.

What the measured numbers actually show

Kimi K3 (from unsloth's repo, shards summed):

Quant	Measured file	VRAM/RAM needed (8K ctx)
UD-Q1_0	466.4 GB	561.1 GB
UD-Q4_K_XL	1.51 TB	~1.6 TB
F16	~3 TB	server farm

So the "594 GB" pages were roughly right for Q1, the "1.51 TB" pages were right
for Q4 — and the people arguing in the comments were both right about different
quants. The site kills the argument by showing the whole ladder.

The honest limitations

Repos without a readable config.json get the flat overhead rule (~20% + 1.5GB). Marked on-page.
MoE models count total size — all experts stay resident in memory even though only some fire per token. This surprises people used to "active params" marketing numbers.
Default context is 8K. KV cache grows linearly; each page explains the math.
CPU inference speed is a different question (memory bandwidth, not capacity) — the FAQ gives rough tok/s bands instead of pretending to precision.

Everything is open source: github.com/Alex20Sas12/modelfit.
If your favorite model is missing or a number looks wrong, the pipeline is
three files — PRs welcome.

Data: Hugging Face public API. Model sizes belong to their uploaders
(unsloth, bartowski, lmstudio-community and others — the community doing the
real quantization work).

I spent a week building 10 Excel templates that actually calculate — post-mortem (and one is free)

Kata Omel — Mon, 07 Sep 2026 16:10:30 +0000

I spent a week building 10 Excel templates that actually calculate things — here's what broke, what worked, and where you can grab one for free

Last week I built a small storefront called SheetVault: Excel templates for budgeting, weddings, crypto portfolios, mortgages, job hunting, savings challenges, freelance invoicing, and splitting bills as a couple. Ten paid templates ($4–8), one free lite version, no email wall on the free one.

This is the honest engineering post-mortem. What broke, what I'd do differently, and — because a budget spreadsheet is only useful if you can actually open it — where to get the free one.

The free one

Free Budget Planner Lite — 3 sheets (Setup, Monthly, Dashboard), real formulas, demo data pre-filled so you can see how it works before deleting the demo rows. Direct download, no signup, no email, no "join 4,000 subscribers first". The source file is also on GitHub: sheetvault-free-templates.

It's the lite version of the $5 full planner, so yes, it's also a funnel. But the funnel works only if the free product is genuinely useful on its own — so it tracks income, fixed/variable expenses, savings rate, and a health indicator. That's a complete monthly budget.

What "actually calculates" means

Most template packs on marketplaces are pretty pictures: hardcoded numbers, static "dashboards" that are just colored cells. Ours use real Excel formulas end to end:

PMT()-based mortgage amortization with extra-payment scenarios (one-time or monthly) and interest-saved-vs-baseline math
Proportional bill splitting by income share, with a settlement sheet that outputs a live verdict: "Jordan owes Alex: $598.80"
Weighted offer comparison: 4 job offers × 5 criteria with customizable weights → automatic WINNER
Every sheet follows the same rule: yellow cells = your input, everything else is a formula. If a user can accidentally overwrite math, the design failed.

Validation was the hard part. I run every file through a formula-evaluation library headlessly (no Excel on the build machine) and diff computed values against independently calculated expectations: PMT checks, savings-rate checks, split totals. A 52-week challenge tracker must show week N × step, projected total step × 1378 — off-by-one errors in week counts are the classic silent killer.

What broke

Data validation lists referencing other sheets must be quoted. 'Settings'!$A$2:$A$9 breaks in some readers if the sheet name contains spaces and you forget the quotes. Grep the raw XML for formula1 entries after every build.
Unicode in base64 payloads. The storefront embeds .xlsx files base64-encoded in a Cloudflare Worker. Using JS btoa() on strings with em-dashes throws InvalidCharacterError — you need TextEncoder first. Classic.
Preview images and file naming drift. One product page 404'd on its cover because the asset was renamed during a refactor. Now every deploy runs a link/asset check over all product pages.

The storefront stack (for the curious)

Cloudflare Workers free tier: server-rendered product pages, JSON-LD Product schema, sitemap, /api/order that generates a unique-cent amount per order (e.g. $5.37, $6.84) so incoming USDT transfers can be matched to orders automatically via Tronscan API
Payments: USDT TRC20 straight to a wallet — non-custodial, no payment processor holding funds, no KYC. Plus a second checkout on sheetvault.sell.app (TRX + ETH USDT, files auto-delivered after payment, 3% fee)
Third checkout for AI agents: every template is also listed on publish.new with an x402 paywall — if an agent (or you, via a 402-aware client) hits the content endpoint, it can buy and download the file directly in USDC on Base, no account needed
Traffic: Pinterest pins (queue-based, 3–6/day), Dev.to/Medium posts like this one, GSC + IndexNow for search

No, it hasn't made money yet — it's been live for about a week. The honest metric right now is: does the free planner get downloaded and does anyone find the paid ones. I'll write the follow-up with real numbers either way.

Takeaways if you're building templates yourself

Ship the free version without an email wall. It converts better into trust (and backlinks) than any lead magnet form.
Validate formulas headlessly. If you can't compute the expected value by hand, you can't test the template.
Unique-cent order matching is a poor man's payment webhook — crude, but it works with zero infrastructure and zero fees on TRC20.

Grab the free Budget Planner Lite, and if you hate it, at least it cost you nothing but 8 KB.

I built a free budget spreadsheet (no signup, no email wall) — here's the whole thing

Kata Omel — Mon, 07 Sep 2026 12:32:34 +0000

I built a free budget spreadsheet (no signup, no email wall) — here's the whole thing

Most budget tools want your email, your phone number, and $8/month before you've even seen what they do. I went the other way: I built a plain .xlsx budget planner that downloads with one click. No signup, no tracking pixel on the download, no "start your free trial."

Here it is:

Download page: https://sheetvault.shablony-pro.workers.dev/free-budget-planner-lite/
Direct file (raw xlsx): https://raw.githubusercontent.com/singap2002-code/sheetvault-free-templates/main/budget-planner-lite_en.xlsx
Source repo: https://github.com/singap2002-code/sheetvault-free-templates

What's inside

Three sheets, real formulas (not pasted values):

Setup — you list your income streams and fixed expenses once. Each row has an amount and a frequency (monthly / weekly / yearly), and the sheet normalizes everything to a monthly number.
Monthly — 12 months across the top. Pulls your Setup numbers automatically, then you fill in variable spending per month. It computes what's left over after everything.
Dashboard — one-glance summary: total monthly income, total fixed costs, average leftover, and a savings-rate readout with a simple traffic-light status (🟢 healthy / 🟡 tight / 🔴 over-extended).

The savings rate is the number I'd argue matters most. If income is irregular, chasing a fixed dollar savings goal is demoralizing — a rate (say, 20% of whatever came in) scales with reality. The Dashboard shows exactly that.

Why a spreadsheet and not an app

It works offline and forever. Nobody sunsets your .xlsx next quarter.
Your financial data stays on your machine.
You can open the formulas and audit them. An app is a black box; a spreadsheet is source code you can read.

Honest limitations

It's the "Lite" version: one currency, no debt payoff schedule, no investment tracking.
If you want those, I sell fuller templates on the same site (budget planner full version, mortgage payoff planner, couple bill-splitter) — that's how I keep the free one free. No pressure, the lite file is genuinely useful on its own.

If you want to tinker

The file is on GitHub (link above), so you can fork it, strip the formulas apart, or rebuild it in Google Sheets. If you make something cool out of it, I'd love to see it.

Questions about the formulas or how the frequency normalization works? Drop a comment — I'll answer.

Want the full version? The complete Monthly Budget Planner (6 sheets, auto dashboard with charts, Plan vs Actual variance tracking) plus 10 more templates — wedding budget, crypto portfolio, mortgage payoff, savings challenge — are in the SheetVault shop: sheetvault.sell.app ($4–$8, instant download, crypto payments welcome). The lite version above stays free forever.

Budgeting apps cost $200-330 over 3 years. Here's the honest spreadsheet math.

Kata Omel — Mon, 07 Sep 2026 06:56:12 +0000

I build spreadsheet tools for a living, so take this with the appropriate grain of salt — but the budgeting-app pricing math of 2026 deserves an honest post.

The 3-year cost table nobody shows you

Tool	2026 price	3-year cost
YNAB	$109/yr	$327
Monarch Money	$99.99/yr	$300
EveryDollar Plus	$79.99/yr	$240
PocketGuard	$71.88/yr	$216
A purpose-built spreadsheet	$5-16 one-time	$5-16

Marketing pages show you the monthly number. The real number is the 3-year one.

Context: Mint — the free app that anchored this whole category — was shut down by Intuit in March 2024. 3.6M users had to migrate, and many discovered their years of transaction history didn't export cleanly anywhere. That's the structural risk of rented software: it can die. A spreadsheet file from 2026 still opens in 2036.

What apps genuinely do better

Being honest, because a one-sided post is useless:

Automatic bank sync — transactions appear without typing. This is the one irreplaceable feature.
Notifications — "you overspent on dining" pings actually work.
Live multi-device sync for households living in different apps all day.

If you know yourself and will never log anything manually — pay for sync. YNAB's method is genuinely excellent and its coaching content is worth something.

What spreadsheets genuinely do better

Ownership. The file can't be paywalled, redesigned, or sunset. YNAB users have lived through two forced migrations.
No credential sharing. Bank-sync apps route your login through aggregators (Plaid & co). Read-only, generally safe — but a third party holds your tokens. A spreadsheet's attack surface is zero.
Custom logic. 50/30/20, envelope, zero-based, income-ratio splitting for couples — change a formula, not a feature request.
Cost. One coffee vs $200-330 over three years.

The decision rule

If you'll maintain manual entries for ~10 minutes a week, the spreadsheet wins on cost, privacy and longevity. If not, buy the app — but budget the 3-year number, not the monthly one.

The behavioral-science footnote: manual entry friction is a feature, not a bug. Studies of budgeting adherence keep finding that people who feel each transaction stick to budgets longer. The apps that removed friction also removed the mindfulness.

Full comparison (with the numbers verified)

I wrote the long version with per-app breakdowns, FAQ, and what Mint refugees actually switched to:

👉 Spreadsheet Budgeting vs Budgeting Apps in 2026

Related honest comparisons from the same series:

Splitwise alternative math — why the free tier got worse and when the app still wins
Mint alternative guide — where the 3.6M users actually went
Streaks vs habit spreadsheets — streak anxiety is real

Disclosure: I sell spreadsheet templates ($5-16 one-time). Every comparison above includes a "when the app wins" section, because dishonest comparison pages are exactly what search engines and readers punish.

What do you use for budgeting in 2026 — app or sheet? Genuinely curious where the balance sits for developers who automate everything else.