Video: "Kimi K3 in 10 Minutes" by Julian Goldie on YouTube.
What Kimi K3 actually is
Kimi K3 is Moonshot AI's third-generation open-source model. At 2.8 trillion total parameters, it is the first open-weight model to reach that scale — larger than anything previously available without a commercial licence. It went live on 16 July 2026 through the kimi.com interface and the Moonshot API, with full open-weight download promised by 27 July.
The architecture is Mixture-of-Experts: 896 expert sub-networks in total, with 16 activated per token during inference. That means the active parameter count per inference is much smaller than 2.8T — the model routes each token through a subset of its capacity rather than running the whole thing. That is how it stays fast enough to be practical rather than just impressive on a spec sheet.
Why the architecture matters: Kimi Delta Attention
The part of Kimi K3 that is genuinely new is Kimi Delta Attention (KDA). Standard attention mechanisms get progressively slower as context grows — which is why 1-million-token windows have historically been painful in practice even when models claim to support them. KDA speeds up decoding by up to 6.3x on long contexts by changing how the model attends to previous tokens.
In practical terms: a task that requires reading a large codebase, a long document history, or a multi-session agent log before responding does not slow down as dramatically as it does on other large-context models. Worth knowing for agent work, where context tends to accumulate across steps.
K3 Max versus K3 Swarm Max
Two variants shipped at launch. K3 Max is the standard one — optimised for single-threaded reasoning, coding, and chat. K3 Swarm Max is built for running many agent tasks in parallel. The Swarm variant co-ordinates multiple sub-agents across a shared context, which is directly relevant for anyone using Hermes Swarm Mode or running multi-agent SEO workflows.
For most starting setups, K3 Max is the right choice. Swarm Max is worth considering once you have a working Hermes or OpenClaw configuration and want to scale it — not before.
Benchmarks: where it actually holds up
On Julian Goldie's Goldie Bench scoring, Kimi K3 lands ahead of GPT-5.5 and roughly level with Opus 4.8 on long-horizon agentic tasks and coding. On SWE-bench style coding evaluations it edges ahead of Claude Sonnet on several runs. On pure chat quality or creative tasks, the gap is less clear.
That said, benchmarks run on specific test sets. Real-world performance depends heavily on the task type, the prompt, and the tools you connect. The result that matters is whether it does your specific job better than what you are running now — not whether it tops a leaderboard compiled by someone else.
What it does not change
Kimi K3 is a model, not an agent setup. You still need Hermes Agent or OpenClaw to give it tools, memory, and the ability to take actions. The model itself cannot browse, write files, call APIs, or manage tasks without an agent layer wrapping it. A bigger brain does not automatically mean better outputs if the scaffolding around it is wrong.
It also has rate limits at the free tier. Basic access at kimi.com will bump into them quickly on multi-step work. For anything beyond simple testing you need a Moonshot API account with a small deposit — roughly $20 gets you to Tier 2, which is enough for regular agent sessions without constant interruptions.
Where this connects to NordSys
NordSys installs and maintains Hermes Agent and Claude Code for UK businesses. Connecting a new model like Kimi K3 to your existing agent setup is a ten-minute configuration change — we handle that as part of the ongoing care plan, usually within a working day of a notable new model release. If you are not yet running a local agent, this is a reasonable moment to start: the open-source model options are genuinely competitive now.
See our AI Agents service →