Kimi K3.1 Pro

Overview

Kimi K3.1 Pro is Moonshot AI's October 23, 2026 update to K3. The upgrade case is stronger task execution and a different cost profile: the one-million-token context ceiling already existed in Kimi K3.

For existing K3 users, standard Pro is worth evaluating first for demanding coding or research. Highspeed is the option to test when response time matters enough to justify higher consumption. This guide uses official documentation and vendor-reported evaluations; it does not claim hands-on testing or independently measured speed.

What Changed

More Capable Long-Horizon Work

Moonshot describes K3.1 Pro as a post-trained K3 with more reinforcement-learning compute. Its internal Kimi Code Bench 3.0 Long-horizon score rises from 55.95% to 61.2%, a gain of 5.25 percentage points. Internal financial-modeling and research scores also improve. These are signals for evaluating multi-step jobs, rather than proof that every coding or research task improves.

The release also emphasizes 3D and parametric modeling. The user-facing 3D Make experience lets people preview and edit models online and download outputs for 3D printing. Treat generated geometry and documents as outputs to inspect before practical use. These examples show the release's direction; they do not mean K3 lacked 3D or financial capabilities.

A Highspeed Choice Alongside Standard Pro

Moonshot says Highspeed maintains the same model capability while producing output about 2–3 times faster. That is a vendor claim about generation speed, not a guarantee that an entire research or coding job finishes 2–3 times sooner. Tool execution and network waits still affect completion time.

Lower Standard API Rates Than K3

The China API platform currently lists lower input, output, and cache-hit rates for standard Pro than for K3. Highspeed charges more for input and output than either standard Pro or K3. Subscription access is a separate decision from API token billing.

Compared With K3

The comparable coding score comes from the same named internal benchmark in the new release article. Do not compare K3's old Terminal-Bench 2.1 score directly with Pro's Terminal-Bench 4.0 score: those are different evaluations.

Pro vs Highspeed: Cost and Access

These are the official China-platform CNY rates, checked October 24, 2026, per one million tokens. They are not converted US-dollar prices or a quote for another billing region.

For equal volumes of uncached input and output, standard Pro's respective unit rates are 35% and 32% below K3. This does not establish total job savings: reasoning length, cache policy, retries, and successful completion all matter. Cache-write TTL affects billing; consult the current pricing documentation before estimating a cached workload.

For Kimi Code membership, Plus / legacy Moderato and above can use standard Pro. Pro / legacy Allegretto and above unlock up to 1M context and Highspeed. Kimi Code describes Highspeed as approximately twice standard Pro's consumption. These are Code entitlements, not a universal promise for every Kimi app feature.

Start with standard Pro for background work or budget-sensitive evaluation. Test Highspeed on interactive work where waiting is costly, comparing total consumption and completion time on the same tasks.

Access and Migration

K3.1 Pro is available through Kimi web/mobile, Kimi Work, Kimi Code, and the Kimi API. Enterprise administrators must enable organizational access, which is off by default at launch. The China API quickstart requires a paid top-up to unlock this flagship; its new-user voucher does not unlock it.

Developers must distinguish the two services:

1. Start a fresh conversation when moving from K3. Kimi warns that changing models inside an existing session can reduce quality and invalidate cached context.
2. Keep thinking enabled. Pro supports low, high, and max, with max as its default; requests that disable thinking are rejected.
3. Use the documented model ID and the endpoint for your service. Kimi Code warns that an incorrect ID can route to a different backend. Preserve the required reasoning history during multi-step tool calls.
4. Evaluate with a fixed effort level and representative tasks. Measure successful completion, generated tokens, latency, and errors before changing a stable application route.

Who Should Upgrade / Who Should Wait

Evaluate Pro now if K3 struggles with multi-step repository work, financial-document synthesis, or iterative 3D creation. Ordinary users can begin with a familiar task and inspect the result; developers can compare matched jobs and actual token use.

Try Highspeed if interactive waiting is the main constraint and your membership or API budget supports higher consumption. Faster generation alone is not a reason to move every background job.

Keep K3 for now if it already meets your quality and latency targets, or if you require downloadable weights today. Moonshot plans Pro's full weights by November 4; availability through hosted services does not establish self-hosting availability or final deployment requirements.

Known Limitations

Moonshot's benchmark results and examples are not independent product tests. Some evaluations are internal, and tool-assisted results depend on their setup. The English launch article and China API guide report different factual-error stress-test figures; this guide does not merge them into a single claimed accuracy rate.

Membership limits and regional API billing differ. Verify your account's actual access and billing before switching. The 1M maximum does not mean every membership has that context allowance.

Sources

• Official K3.1 Pro launch article: model changes, vendor evaluations, product access, and planned weights.
• Kimi Code release notes: October 23 release date.
• Kimi Code model configuration: membership access, model IDs, consumption, and switching behavior.
• K3.1 Pro API quickstart: general API model IDs, access and reasoning controls.
• Kimi API platform: current China-platform model rates.
• API pricing and billing: token billing and cache-policy distinctions.
• Official K3 model card: existing context ceiling and released weights.