Why this release is different
Open-weight models have been “almost frontier” for a while. The interesting thing about K3 is how small the “almost” has become, and how the gap closed. The scores first, with the usual self-reported-benchmark salt. On Artificial Analysis’s Intelligence Index, K3 lands third overall — behind only Claude Fable 5 and GPT-5.6 Sol at maximum effort, and ahead of everything else, closed or open. On Artificial Analysis’s private long-horizon agentic evaluation it reaches an Elo of 1547 — a 732-point jump over its predecessor K2.6. It takes first place outright on the Frontend Code Arena, ahead of the closed frontier. And on GDPval-AA v2, a benchmark built from real tasks across 44 occupations, it scores 1,687 — third again, ahead of Claude Opus 4.8. A model you can download outperforming last generation’s closed flagships and trailing only the current two — that has never happened at this margin before. Nathan Lambert’s read is that the open-to-closed gap has compressed from six-to-nine months to more like three to five, and K3 is the proof point. The second reason it matters is how Moonshot got there: architecture, not just scale. K3 activates only 16 of 896 experts per token — an unusually sparse MoE — and pairs it with two new mechanisms, Kimi Delta Attention (a hybrid linear attention that makes the 1M-token window economical) and Attention Residuals (which carry representations across model depth). Moonshot claims roughly a 2.5× overall scaling-efficiency improvement over the K2 generation from these choices. The design is agent-shaped from the ground up: 1M tokens of context, aggressive prefix caching (cache-hit input is **3.00 uncached), and long-horizon agent training. The bet is that agents get cheap when the architecture, the serving stack, and the agent loop are designed together — which happens to be exactly the shape of workload a canvas-driving ComfyUI agent produces: a long, stable tool-schema prefix and many short turns against it.The honest caveats
We wired K3 in, used it, and read the fine print, so here is the other column. It thinks. A lot. Always. K3 launched with a single reasoning effort — max — and no way to turn it down (Moonshot says lower-effort variants are coming). Simon Willison measured a run where 13,241 of 16,658 total output tokens were reasoning. At **0.95/$4 — asking K3 to nudge one widget value is using a freight train to deliver a postcard. The mitigations are real, though: K3 uses about 21% fewer output tokens than K2.6 for equivalent work, and the >90% cache-hit rates Moonshot reports on coding workloads map well onto agent loops like ours. Moonshot’s own launch notes are unusually candid: overall performance still trails Claude Fable 5 and GPT-5.6 Sol, the model is sensitive to having its thinking history dropped by a harness, and it shows “excessive proactiveness” on ambiguous tasks — it will confidently do something when a more conservative model would ask. Supervise accordingly; every canvas edit the agent makes is one Ctrl+Z from gone, which is precisely the safety net an eager model needs. And the headline is a promise until July 27. “Open weights in nine days” is not yet open weights. Moonshot’s track record here is good — K2’s weights shipped as promised — but the strategic significance everyone is writing about is conditional on the Hugging Face upload actually landing.Wired in, two days later
comfyui-mcp already speaks to fourteen backends through one orchestrator port, so a new OpenAI-compatible provider is a well-worn path. K3 landed as themoonshot provider in PR #233,
shipped in v0.38.0 on July 17 — the day after K3’s release — and
live-verified against the real API that night: our listModels() sees
kimi-k3 in Moonshot’s /v1/models, and a real turn answers over
chat/completions. The panel’s Kimi K3 chip followed in panel 0.9.1 the
next day.
One naming knot worth untangling, because you’ll see two Kimi chips in the
panel and that’s deliberate. Kimi K3 lives on Moonshot’s platform API
(api.moonshot.ai/v1, key from platform.kimi.ai,
pay per token). The pre-existing Kimi chip is Kimi Code — the separate
coding-subscription product with its own OAuth login and endpoint. Same
company, different products, different auth, both selectable side by side.
Setup is the whole of this paragraph: get a key from platform.kimi.ai, paste
it into the Kimi K3 (Moonshot) slot in the panel’s API Keys card (or set
MOONSHOT_API_KEY for MCP-only use), pick the chip, Connect. Default model is
kimi-k3; COMFYUI_MCP_MOONSHOT_MODEL and COMFYUI_MCP_MOONSHOT_BASE_URL
override the model and endpoint — the latter means that when the weights drop
on the 27th, pointing the same provider at a self-hosted or third-party K3
endpoint is a one-variable change. That’s the quiet payoff of the open-weights
part: the provider you configure today against Moonshot’s API is the same
provider that runs against your own serving stack later.
Two integration notes from the trenches. K3 streams its always-on thinking as
reasoning_content alongside the answer; the backend handles it correctly
today, and surfacing the live reasoning stream in the panel is on the wish
list. And K3 was also the straw that fixed a maintenance problem: it was the
provider that made us count the ~10 scattered files every new
OpenAI-compatible backend had to touch, and the follow-up
provider registry collapsed
that to one table entry — the next K3-shaped release wires in faster than this
one did.
Vision works the way it works everywhere in comfyui-mcp: capability is judged
per model, not per provider, so K3’s native visual
understanding means it can actually look at your renders when you ask “why is
the face melted?” — delivery is attempted, and honesty is enforced if a model
can’t see.
Where it fits in your lineup
The panel’s whole thesis is pick a provider, not a port, and K3 slots into a real niche in that lineup. For quick widget nudges, a free local model on your own GPU is still the right tool — K3’s always-max thinking is overkill you pay for by the token. For the absolute ceiling, Claude and GPT-5.6 still hold the top two slots on most boards. K3’s lane is the long, gnarly middle: a 128-node workflow autopsy, a plan-install-run-verify chain that touches thirty tools, anything where the 1M-token window and long-horizon training earn their keep — at cached-input prices that undercut the closed frontier on exactly those workloads. And it’s the one model in that class you’ll be able to download. nine days until the weights. The chip is already in your sidebar.Try it: install comfyui-mcp and add the Panel to ComfyUI, grab a key at platform.kimi.ai, and pick the Kimi K3 chip. If you get K3 doing something great on your canvas — or something hilariously overeager — file an issue or share it; we read everything.