On October 23, 2026, Vercel put Moonshot AI’s Kimi K3.1 Pro on its AI Gateway alongside a faster Highspeed variant, routable from Claude Code, Codex, and Cursor. On November 4, Moonshot says it will release the model weights. That date changes the shape of the decision in front of you: routing traffic through the gateway is a reversible trial you can start today at provider pricing with no markup, while self-hosting is a capital commitment gated on two things nobody has seen yet, the weights themselves and the license attached to them. The practical answer, given what is actually documented, is to route now, evaluate on your own workloads, and defer any self-hosting spend until the weights exist as an artifact rather than a plan.
Everything below separates what Moonshot and Vercel have stated from what anyone has independently measured. That distinction matters more than usual here, because the changelog announcing the release cites no third-party evaluation at all.
What shipped on October 23
Vercel’s changelog announces two model IDs: moonshotai/kimi-k3.1-pro and moonshotai/kimi-k3.1-pro-highspeed, the latter described as “a faster version for interactive use.” Per the changelog, “K3.1 Pro is the first model in Moonshot’s K3.1 series. It builds on Kimi K3 with improvements in coding, 3D modeling, financial analysis, and agentic tasks, and keeps the 1M-token context window and image input.” Those improvement claims are vendor-reported; the changelog cites no third-party evaluation, and every capability claim it carries comes from Moonshot or Vercel themselves.
The routing path is short. Install the latest Vercel CLI, run npx vercel ai-gateway setup, then select either model ID inside Claude Code, Codex, or Cursor. Both variants are also callable through the AI SDK and the Chat Completions, Responses, and Anthropic Messages APIs, so direct integration without a coding agent is possible too.
The pricing mechanism is what makes this a trial rather than a commitment. The changelog states: “AI Gateway reflects provider pricing with no markup and does not charge a platform fee on inference, including on Bring Your Own Key (BYOK) requests.” No markup is not no cost. You still pay Moonshot’s token rates, and BYOK still runs through your existing provider account. But there is no gateway tax, no migration of your agent setup, and no lock-in: switching the model ID back is a config change.
Pro versus Highspeed: pick by session shape, not by name
The two variants split along one axis: how long the model needs to think before you need it back.
Vercel positions K3.1 Pro for sustained work: “For coding, K3.1 Pro handles work ranging from small fixes in an existing codebase to building a full project, including long agent runs with many steps.” That is the workload profile of an unattended Claude Code or Codex session grinding through a multi-file refactor, where per-token latency matters less than whether the run completes coherently. Highspeed is positioned only as “a faster version for interactive use,” which maps to the pair-programming loop in Cursor or a supervised session where you are waiting on each response.
One honest limitation: the latency and quality tradeoff between the two variants is quantified nowhere in the changelog. “Faster” could mean a distilled variant with degraded reasoning, a serving-tier difference with identical weights, or something else; the changelog does not say. Treat the Pro/Highspeed split as a starting heuristic, not a measured finding, and test both on a representative task if interactive latency is what you actually care about.
There is a second reason to be careful with long agent runs specifically, and it comes from Moonshot’s own prior work. The Kimi K2 technical report documented failure modes in the previous generation: “When dealing with hard reasoning tasks or unclear tool definition, the model may generate excessive tokens, sometimes leading to truncated outputs or incomplete tool calls.” K3.1 Pro is several releases later and these specific issues may be fixed, but nothing in the Vercel announcement demonstrates that for the new model. If your workload is exactly the long multi-step agent run the Pro variant is pitched at, your own eval is the only evidence you will have before November 4.
Vendor claims versus measured evidence
Here is the state of the evidence, stated plainly:
- Claimed, vendor-reported: improvements over K3 in coding, 3D modeling, financial analysis, and agentic tasks; “fewer factual errors than earlier Kimi models on difficult prompts”; a 1M-token context window; image input.
- Measured independently for K3.1 Pro: nothing in the Vercel announcement.
- Measured for the predecessor K3: the K3 technical report concedes the model “trails the strongest proprietary systems overall—Claude Fable 5 and GPT-5.6 Sol—” while leading the other open and proprietary models in Moonshot’s suite. Vendor-cited third-party standings as of July 23, 2026, put K3 second of 39 on the Vals Index (74.7%, behind Claude Fable 5 at 75.1%) per the K3 report, and first of 99 on WebDev Arena at 1,678 Elo. On the Artificial Analysis Intelligence Index v4.1, the same report places K3 fourth of 580 at 57.1, behind Claude Fable 5 (59.9) and GPT-5.6 Sol (58.9); it counts as third only if GPT-5.6 Sol effort variants are treated as a single entry.
Those K3 numbers describe the model K3.1 Pro builds on, not K3.1 Pro itself. They establish that Moonshot ships competitive open models; they say nothing about whether the new release closes or widens the gap to the proprietary frontier. If your routing decision depends on beating Claude on your tasks, the only honest answer today is: unknown, test it.
The scale claims inherited from the K3 generation are at least well documented. K3 is a 2.8-trillion-parameter Mixture-of-Experts model with 104 billion activated parameters, native vision, and a 1M-token context, per its technical report, and Moonshot’s homepage repeats the 2.8T-parameter and 1M-token-context figures for the K3 generation. The architecture traces to Kimi Linear, Moonshot’s hybrid linear-attention design, which the authors report outperforms full attention under fair comparisons across short-context, long-context, and reinforcement-learning regimes. That architecture detail matters later, when we get to self-hosting: linear-attention variants have different inference memory behavior than standard transformers, which affects what serving the model locally would actually look like.
November 4 is a plan, not an artifact
The changelog’s exact words: “Moonshot AI plans to release the model weights by November 4: support for more providers will follow.” Plans slip. The question is how much weight to give this one.
Moonshot’s track record is genuinely good on this axis. K3 launched on July 16, 2026, the first open model to reach 2.8 trillion parameters, with a custom license attached, per Wikipedia’s Moonshot AI entry. The K3 report itself states: “We release the full Kimi K3 model weights, making frontier intelligence available for research, deployment, and further innovation.” Earlier, the K2 base and post-trained checkpoints went public on HuggingFace, and the K2.5 release continued the pattern with an open post-trained multimodal agentic checkpoint. A vendor that has shipped weights for three consecutive releases deserves more credit for a dated commitment than one making its first promise.
Credit is not a guarantee, and the sensible posture treats the window before November 4 as one with a checkpoint at the end. If weights appear on or near November 4, the self-host evaluation begins. If the date slips without a revised commitment, that tells you something about how Moonshot prioritizes this release, and the gateway trial keeps running meanwhile at no strategic cost.
Self-hosting is a license question wearing a hardware question’s clothes
Open weights does not mean permissive terms, and K3’s precedent is the cautionary exhibit. “Despite its open weights, K3’s custom license requires revenue sharing of up to 30% for inference providers generating over US$20 million annually,” per Wikipedia’s Moonshot AI entry. If you are an individual developer or an internal platform team, that clause likely never touches you. If you are building an inference product on top of Kimi weights, it restructures your unit economics entirely. K3.1 Pro’s license terms are unstated in the changelog; assuming they match K3’s, or that they will be more permissive, is speculation in both directions.
The hardware side is similarly unresolved. No K3.1 Pro hardware requirements appear in the changelog. Extrapolating from K3’s 2.8T-parameter MoE scale suggests a serious multi-GPU serving footprint, but that is inference from the predecessor, not evidence about the new model, and MoE activation patterns plus the linear-attention architecture make naive parameter-count math unreliable. The disciplined move is to wait for the release artifact and read its actual deployment notes.
This is why the two decisions should be held separately. Routing through the gateway commits you to nothing beyond token spend. Self-hosting commits you to hardware, ops, and whatever license Moonshot attaches, none of which can be priced until November 4 at the earliest.
The compliance backdrop is part of the routing decision
Two documented facts belong in a procurement conversation even if they do not change an individual developer’s weekend experiment. Throughout 2026, Anthropic accused Moonshot of distilling Claude and covertly routing requests to it, and from April 2026 a US congressional committee pressed American companies over their use of Kimi and other Chinese AI models, per Wikipedia. These are allegations and political scrutiny, not adjudicated findings. But if you work somewhere with vendor-review processes, government contracts, or data-residency constraints, routing proprietary code through a Chinese frontier model is a decision your security team will want on record. Gateway routing makes the trial technically easy; it does not make it institutionally approved.
A dated decision framework
| Checkpoint | What to do | What it tells you |
|---|---|---|
| Now through Nov 3 | Route a slice of Claude Code/Codex/Cursor traffic to moonshotai/kimi-k3.1-pro; use Highspeed for interactive sessions; run your own task suite | Whether vendor capability claims hold on your code, at zero markup |
| Nov 4 | Check whether weights shipped, and read the license before anything else | Whether the self-host path exists and what it costs legally |
| Post-release, if weights ship | Price hardware against the actual deployment notes; compare gateway token spend versus serving cost | Whether self-hosting beats continued routing on economics |
| If the date slips | Keep or drop the gateway trial on its measured merits; do not pre-commit self-host budget | The vendor’s stated plan failed; the trial evidence still stands |
The column that matters most is the second one. Every step is reversible until hardware gets purchased.
Verdict: route now, and know what would change that
I would route a bounded slice of agent traffic to moonshotai/kimi-k3.1-pro this week, keep Highspeed for interactive sessions, and treat November 4 as a decision checkpoint rather than a countdown. The conditions where that recommendation changes: if your evals show the model truncating outputs or dropping tool calls on your hard tasks the way K2’s report documented for its generation, drop it and lose nothing but a few dollars of tokens; if the weights date slips, the self-host half of the plan is void but the trial is unaffected; if the shipped license carries K3-style revenue sharing and you are anywhere near being an inference provider, the self-host economics need a lawyer before they need a GPU quote.
What this article cannot tell you is whether K3.1 Pro is good. Every capability claim for it comes from Moonshot or Vercel, the Highspeed tradeoff is unquantified, and the changelog cites no independent evaluation. The gateway trial exists precisely to close that gap on your own terms, before the commitment half of the decision even becomes available.
Frequently Asked Questions
What are the license terms for self-hosting Kimi K3.1 Pro?
K3.1 Pro’s license terms are unstated in the changelog; assuming they match K3’s, or that they will be more permissive, is speculation in both directions.
Join the discussion
Share a useful perspective or ask a question about this article.