Kimi K3 and the open-model wave
Kimi K3 is a 2.8T open-weight model ranked fourth at the frontier. BYOK developers: when to pick it in Cortex vs GLM 5.2 or DeepSeek.
The short answer: Kimi K3 is a 2.8-trillion-parameter open-weight model that independent tests place fourth at the frontier. This page is for BYOK developers who want whole-codebase reasoning without renting one closed lab. Use K3 for long-context work. Keep DeepSeek or Qwen for cheap loops. Do not treat it as a wholesale switch.
Who is this Kimi K3 guide for?
Developers who already bring their own API key and need a model that can hold a whole repository in context. Read this if you run Cortex, you pay OpenRouter or Moonshot directly, and you are tired of one vendor deprecating the model you just learned. Skip it if you only want inline autocomplete; K3 is built for agentic loops, not tab-complete.
On July 16, 2026, Moonshot AI released Kimi K3. Independent testing places it fourth among all frontier models, behind Claude Fable 5 and GPT-5.6 Sol, and just ahead of Claude Opus 4.8. The ranking is interesting. The license is the point. K3 is an open-weight release, the largest ever, with full weights expected by July 27. Frontier-class intelligence is no longer something you only rent from a closed lab. For BYOK developers, that changes the calculation.
What is Kimi K3?
Kimi K3 is a 2.8-trillion-parameter mixture-of-experts model with a 1-million-token context window, native vision, and open weights. Of its 896 experts, only 16 (about 1.8%) activate per token, so inference costs a fraction of the headline parameter count.
- 2.8 trillion parameters, mixture-of-experts. Sparse activation keeps the bill closer to a mid-size model than a dense 2.8T giant.
- 1-million-token context window. Whole-codebase scale, not file scale. That is the difference between asking "where is auth?" and pasting three files by hand.
- Native multimodal. It reads screenshots, diagrams, and UI mockups, which matters when the agent has to match a design or a failing window.
- Always-on thinking mode. Built for the deliberate reasoning that agentic loops demand.
- API pricing: $3 input / $15 output per million tokens. Claude-Sonnet-class pricing for top-four-frontier performance.
We added K3 in Cortex v2.8.6, two days after the public release, through OpenRouter. On our test machine we ran a 40-file Django auth refactor against the 1M window: the model kept the token-refresh path and the license check in the same pass, which smaller context windows split into two chats. That is the original-experience reason we care about this model, not the press ranking.
Should I switch from Claude or GPT to Kimi K3?
No. Route by job. Do not migrate the whole stack. Closed labs still win some tasks. Open weights win permanence and price. Cortex is built so you can pick per task.
| If you need | Use this first | Leave for later |
|---|---|---|
| Whole-repo reasoning, 1M context, screenshots | Kimi K3 | A closed model you already pay for by habit |
| Leading open-weight coding model on evals | GLM 5.2 | K3 when you only need a short function edit |
| Bulk iterations, cheap loops | DeepSeek or Qwen | K3 at $15/M output for every retry |
| A specific closed-lab behavior | Claude or GPT, your key, your call | Treating one vendor as the only IDE |
That mix is the BYOK play. You keep one IDE. You change the dropdown. You do not migrate tools when next month's open-weight winner ships.
How do I use Kimi K3 in Cortex?
Add your OpenRouter key in Settings, then pick Kimi K3 from the model dropdown. Full 1M context and vision are included. No extra plugin.
- Open Settings (gear at the bottom of the left sidebar) and add an OpenRouter API key. Keys stay in your operating system's secure credential store, not in a plaintext file.
- Pick Kimi K3 from the model list. Cortex added it in v2.8.6, two days after Moonshot's release.
- Use it on tasks that need the 1M window or a screenshot. Route cheaper models for bulk retries so the $15/M output line does not eat the budget.
Because Cortex is BYOK with a verified loop around whatever model you choose, the practical play is routing: K3 or GLM 5.2 for heavy reasoning and big-context work, DeepSeek or Qwen for bulk iterations, a closed frontier model when you specifically want it. All in one IDE. All at provider list prices. When the open-model wave produces next month's winner, using it is a dropdown change, not a tool migration. See real BYOK prices if you want the bill broken down.
Why does the open-model wave matter?
K3 is the crest of a 2026 pattern: open-weight models closing on the frontier, while closed labs consolidate. Zhipu's GLM 5.2 still leads most open-weight coding evals. DeepSeek and Qwen keep shipping near-frontier quality at commodity prices. Now the largest model ever opened lands fourth overall. Meanwhile the closed side of the industry spent the year consolidating into conglomerates.
Why open weights matter to a working developer, in one sentence each:
- They cannot be taken away. A closed model can be deprecated, repriced, or acquired out from under you. Open weights, once released, are permanent infrastructure.
- They create price gravity. Every open release near the frontier pulls the whole market's API prices down. That is why agentic coding gets cheaper every quarter.
- They enable choice. Multiple hosts serve the same weights, so you pick on price and latency instead of vendor loyalty.
Sources
Frequently asked questions
What is Kimi K3?
Kimi K3 is Moonshot AI's 2.8-trillion-parameter mixture-of-experts model, released July 16, 2026. Independent tests place it fourth among frontier models. It has a 1-million-token context window, native vision, and open weights.
Who should use Kimi K3 for coding?
Developers who already bring their own API key and need whole-codebase reasoning without locking into one closed lab. Skip it if you only want inline autocomplete.
Should I switch from Claude or GPT to Kimi K3?
Not as a wholesale switch. Use K3 or GLM 5.2 for long-context reasoning, DeepSeek or Qwen for bulk iterations, and a closed frontier model when you specifically want it. Cortex is built for that mix.
How do I use Kimi K3 in Cortex?
Cortex added Kimi K3 in v2.8.6 through OpenRouter. Add your OpenRouter key in Settings, then pick Kimi K3 from the model dropdown. Full 1M context and vision are included.