# Marcus Delgado, Senior Software Engineer at Cartwright Health — read of Kimi K2 Code IDE, June 13 2026

> 9 years writing backend services, currently at a 200-person health-tech company where "no code leaves the building" is an actual policy, not a vibe. I have a 7-year-old who wakes up at 6am so my side-project window is 4:45am to 5:55am, and I waste none of it.

## How I got here

Indie Hackers weekly digest mentioned a "local AI coding tool with no subscription" and I clicked expecting a product landing page. What I got was something different and it took me a full scroll to figure out what was being sold. A friend also builds dev tools and I've been watching the local-inference space since Ollama started getting traction. I was genuinely looking, not browsing.

## What I clicked first

The spec table grabbed me. "Zero cloud dependency, zero telemetry, zero subscription creep." That's three things I've complained about out loud in the last month. In healthcare we can't pipe source to a cloud API, full stop. So the framing was correct. Then I hit "80% Token Efficiency vs Claude" and I slowed down because that's a specific claim and I wanted to know what it means. It doesn't say. Smaller model, same outputs is not a number. That sentence is doing a lot of lifting with no support underneath it.

## Where I paused

"$-24,400 Year-1 take-home (Fermi)" stopped me cold. That's a negative number. They are telling me the expected return in year one is losing twenty-four grand. And they left it up there. That's either the most honest thing I've read on a product page this year, or it's a sophisticated way to manage liability while still cashing the $99. I genuinely don't know which. I sat on that for a couple minutes.

## What I distrusted

A few things, in order:

First: "Kimi K2.7-Code runs on your machine." Kimi K2 is Moonshot AI's model. I know this because I was reading about it last week. So is this a wrapper around an existing open-source model, a fork, a rebrand? The page doesn't say. "Community Model, Commercial Support" is not an answer. If you're building on top of someone else's weights, say so clearly.

Second: "One-line setup" for fine-tuning on a company codebase. I have shipped adapter-layer tooling. It is never one line. That phrase is in the spec table like it's a feature but it's a marketing aspiration. Anyone who has actually set up LoRA adapters on a private codebase is going to roll their eyes here.

Third: "we don't have live customers on this idea yet. We shipped the strategy package." So I'm being asked to pay $99 to $199 for a build plan for a product that the people selling me the plan have not proven out. That's the Wishdeal Studio model, I get it. But combined with the -$24,400 Fermi estimate, I need to understand what I'm buying more clearly.

## What would convince me

Not a case study. A technical README. Show me the actual model card for "Kimi K2.7-Code" and whether it's the Moonshot weights or a derived fine-tune. Show me a benchmark head-to-head: Claude Sonnet completing a real refactor task vs this model, token counts side by side, output quality graded by a human. One real example with real numbers would do more than this whole spec table. Also: one paying customer of any Wishdeal Studio package who shipped something. Not a testimonial quote. A link to the thing they built.

## What I'd ask in an email reply

1. Is Kimi K2.7-Code the Moonshot AI model, a fork, or your own training run? What exactly am I getting if I adopt this build?

2. The Fermi math shows -$24,400 in year one. What's the model where someone comes out ahead, and what does that person look like specifically: solo dev, small agency, existing dev tool company?

3. Has anyone actually bought the $99 adopt tier and shipped even an MVP? If yes, can I talk to them for 15 minutes before I buy?

## Verdict: on-the-fence

The honesty about failure odds and negative year-one returns is unusual enough that I haven't closed the tab. But I cannot tell from this page whether the underlying model is real, whether "one-line fine-tune" is literal or aspirational, or whether $99 gets me something I couldn't assemble myself from existing open-source projects in a weekend. The concept is right, the execution of the page leaves too many technical questions unanswered.

---
*Memo by skeptic persona, generated 2026-06-13. Studio breaks own self-grading loop.*