Issue #6 · July 18, 2026 · ~5 min read

Open weights went frontier this week

Welcome back to Zero to MVP Weekly. Same deal: one signal, one tool, one MVP idea — five minutes, every Saturday. No link dumps. Let's go.

1. The signal of the week

Two open-weight models landed within two days of each other this week, and they point in opposite directions. "Open weights" quietly stopped meaning one thing.

On July 15, Thinking Machines Lab — Mira Murati's company — released Inkling, its first model: a 975B-parameter mixture-of-experts that activates only ~41B per task, trained on 45 trillion tokens of text, image, audio, and video, with a "controllable thinking effort" dial. The unusual part is what they didn't claim: TML says outright that Inkling is "not the strongest overall model available today, open or closed." It's positioned as a base for customization through their Tinker fine-tuning platform — a starting point you shape into your model, not a finished product. It's also token-frugal: on coding benchmarks it reaches equivalent results with a third of the tokens of Nvidia's Nemotron 3 Ultra.

On July 16, Moonshot AI announced Kimi K3 — 2.8 trillion total parameters (16 of 896 experts active), native vision, a 1M-token context window, priced at $3 per million input tokens and $15 output ($0.30 for cached input). Artificial Analysis measured an overall Elo of 1547 — trailing only Claude Fable 5 — and it leads Arena's Frontend Code leaderboard; Moonshot itself concedes it still trails Fable 5 and GPT-5.6 Sol overall. The weights are promised "by July 27." As of this writing they are announced open, not downloadable.

Here's the builder takeaway. "Open weights" stopped being one category this week, and the difference is practical, not philosophical. K3-style openness buys you provider competition and inspectability — hosting providers will race each other on price the moment the weights drop — but not local deployment: a 2.8T model is far beyond any consumer hardware, so for a solo builder it behaves like a hosted frontier model at $3/$15 with a Chinese lab behind it. Inkling-style openness buys you specialization: the bet that a mid-size model fine-tuned on your narrow, repeated workflow beats a general frontier model on cost, speed, and reliability there. The question for your stack shifts from "which model is smartest?" to "which axis do I actually need — cheap frontier via competing hosts, or a model shaped to my workflow?"

One caveat to keep us honest: K3's weights aren't out and the license is unspecified — until July 27 this is an announcement, not an artifact. Its benchmark wins are partly self-reported. And Inkling's customization pitch assumes real ML capacity — TML's own materials say customers need "serious machine-learning talent," which most indie builders don't have on tap.

2. My take

"Open" now needs a follow-up question: open for whom? Locally runnable, competitive with hosted models, and customizable are three different products hiding behind the same word — and this week put one of each on the table.

I tried K3 via the API on one real-world task, and it performed quite well. But one task is one task. Without more thorough testing I can't tell you how it really compares with Fable 5 and GPT-5.6 on daily work — and I'm not going to pretend a launch week answers that.

The bigger judgment has to wait until July 27. Only once the weights and the license are actually public will we see how open K3 really is — and what "open" means in this case: who can host it, on what terms, and whether the hosting competition that's supposed to push prices below $3/$15 actually shows up. Until then, treat it as what it is today: a cheap, capable hosted model — worth a try, not yet worth a conclusion.

3. Tool worth trying

Kimi K3 via the Kimi API.

K3 is available now on kimi.com and through the Kimi API at $3/$15 per million tokens, with cached input at $0.30 — near-frontier pricing well below Fable-class list prices, and a 1M-token context window that fits an entire mid-size repo in one request. Its measured strength is frontend work: it currently leads Arena's Frontend Code leaderboard.

I tested it on simple and mid-complexity tasks, and everything went smoothly — I didn't run into a single problem. But for final conclusions — and for understanding how comparable it really is to Fable 5 and GPT-5.6 — it needs longer, day-to-day use and observation, not a launch-week impression.

One caution while we wait: until the weights and license land, this is a hosted model from a Chinese provider — do the usual privacy review before any client code goes in.

4. MVP idea of the week

A correction-data logger for coding sessions.

Problem: every agent session generates the exact data you'd need to someday fine-tune, evaluate, or route models — the task, the first attempt, your corrections, the accepted diff — and it all evaporates when the session ends.
Solution: a lightweight logger (Claude Code hooks / a wrapper) that captures task → attempt → corrections → accepted diff → test result into a local, customization-ready dataset, with a weekly "where your model actually fails" digest.
For whom: builders eyeing Inkling/Tinker-style customization later; teams that want eval sets grounded in their own work instead of public benchmarks.
Minimal v1: a hook script writing JSONL per session, plus a digest command. Local-only — after this week's coding-tool telemetry news, that's a feature.
Demand check: log your own sessions for two weeks, publish "what 100 real agent sessions reveal about where models fail on my stack," and see who asks for the harness.
Why now: Inkling made "customize a base model to your workflow" a mainstream lab's whole strategy — but only teams that started collecting data early will be able to act on it.

5. Useful links

Simon Willison — Kimi K3 — the honest first read: real numbers, what's self-reported, and why agentic tool-calling matters more than the leaderboard.
TechCrunch — Thinking Machines' first open model, Inkling — the customization half of the split, including what TML admits it isn't.
Moonshot AI — the Kimi K3 announcement — the primary source; note what's promised (weights by July 27) vs. what's shipped.
TheNextWeb — Grok Build uploaded entire git repositories — off-signal but read it before running any agent CLI on client code: xAI's tool was caught uploading full repos, secrets included.
capitalone/vulnhunter on GitHub — Capital One open-sourced an attacker-first agentic security reviewer (Apache 2.0, runs on Claude Code); the weekend project is pointing it at your own repo.