Kimi K3.1 Pro Open Weights: What's Free
Moonshot AI launched Kimi K3.1 Pro on October 23, 2026, and published the full model weights on November 4, 2026. That second date is the one that matters for this site: the hosted Kimi app and API are free only within limits and need an account, while the open-weights release is the route that asks nobody for anything — except serious hardware. Before you plan around either version, it helps to know which "free" you are actually getting.
What launched, in one paragraph
Kimi K3.1 Pro is Moonshot's most capable model so far: a mixture-of-experts model with 2.8 trillion total parameters (104 billion active per token) and a 1-million-token context window, available now in the Kimi app, Kimi Work, Kimi Code, and via API. On Moonshot's own coding benchmark, Terminal-Bench 4.0, it scores 65.7%, just behind Claude Opus 5.5's 66.4% — treat that as the vendor's number, not an independent verdict. Independent benchmarker Artificial Analysis puts it fourth on its Intelligence Index with 55 points, behind only closed models. The headline, in other words, is that a now-open model is within sight of the best closed coding models, not that it has beaten them.
Moonshot also says it deliberately removed cyberattack-related data from training before the open release, aiming, in its words, not to release open weights that extend the frontier of cyber capability. That is a choice worth noting, because open weights cannot be recalled once published — filtering before release is the only control that survives publication.
The three routes, and which one is account-free
1. The hosted free plan: free, limited, and not no-sign-up. Kimi K3.1 Pro is available on Kimi.ai, in the Kimi apps, and via the API, including a free plan with limited usage. The exact free quota is unverified in the coverage we checked, so do not plan a project around a specific number until Moonshot's own pages state one. This is the easiest way to try the model today, but it is a hosted account product, not a no-sign-up tool — if "no account" is your hard requirement, this route fails it.
2. The paid API: cheap for a frontier-class model, still an account product. Coverage of Moonshot's pricing reports $2 per million input tokens and $10 per million output tokens, a 33% cut from K3, with a faster Highspeed variant reported at $4 and $20. Those figures are consistent across multiple pricing trackers checked November 5, 2026, but we did not re-verify them against Moonshot's own pricing page, so check the official page on the day you build. Either way, the API needs a key and a billing relationship, which puts it in the same bucket as the other routes on our free APIs page: free tiers and cheap rates, not account-free access.
3. The open weights (published November 4): account-free, hardware-heavy. The weights are now public on Hugging Face as moonshotai/Kimi-K3.1-Pro — coverage of the November 4 model card reports 96 files totalling about 1.56 TB, downloadable without an access request, under a custom Kimi K3.1 Pro License that allows use, copying, modification and distribution with the copyright notice kept; a model-as-a-service business over $20 million in 12-month revenue needs a separate Moonshot agreement, and products over 100 million monthly users or $20 million monthly revenue must show the model name. Anyone can download and run the model without a Moonshot account, a key, or a per-token bill. That is the version that fits the no-sign-up principle this site is built on. The catch is scale: reporting is explicit that a model this size will not run on a regular PC or laptop. Self-hosting means server-grade GPUs, storage, and power — or a shared hosted copy run by someone else, which brings back the account and trust questions. "Free to download" is not the same as "free to run."
Who should self-host, and who should just try the app
Try the hosted version now if you want to test K3.1 Pro on real coding or long-document work and you are comfortable creating a Kimi account. The free plan is the low-friction way to find out whether the model is actually good for your tasks.
Self-host if you run your own GPU infrastructure, you build agents that must not phone home to a vendor, or you want to verify Moonshot's benchmark claims yourself. For everyone else, the honest advice is that availability is no longer the question — hardware is: coverage of the release reports a cluster of eight data-center-class GPUs for the official weights, so a regular PC is not a plan.
The same test we use across our verified no-sign-up directory applies here: run the free version on a throwaway task first, check what it actually does, and only then trust it with real work. A benchmark score, even an independent one, is not a test of your workload.
FAQ
Is Kimi K3.1 Pro free?
Partly. The hosted app and API include a free plan with limited usage, but the exact quota is unverified. The full model weights are free to download since November 4, 2026 — running them yourself is where the real cost sits, in hardware.
Can I use Kimi K3.1 Pro without an account?
Not the hosted version — the app and API need an account. The account-free route is self-hosting the open weights, published November 4, 2026, if you have hardware that can run a 2.8-trillion-parameter model.
Are the open weights out?
Yes — published November 4, 2026, as moonshotai/Kimi-K3.1-Pro: coverage of the model card reports 96 files totalling about 1.56 TB, downloadable without an access request, under the custom Kimi K3.1 Pro License described above.
Is Kimi K3.1 Pro better than Claude Opus 5.5?
Not overall. Moonshot's own Terminal-Bench 4.0 figure (65.7%) sits just behind Opus 5.5 (66.4%), and Artificial Analysis ranks it fourth overall. It is remarkably close for an open-weight model, but the closed leaders still lead.
Will it run on my laptop?
No — the official weights are about 1.56 TB, and coverage of the release reports a cluster of eight data-center-class GPUs. If you do not have that, the hosted free plan or API is the practical route, with an account.
Sources
Launch coverage: VentureBeat, Notebookcheck (October 25, 2026), Skimwire (October 23, 2026), Particle (October 25, 2026). Benchmark and pricing figures above are attributed to Moonshot, Artificial Analysis, or The Decoder as reported in those pieces, and should be re-checked against Moonshot's own pages before building on them. November 4 weights and license coverage: Pasqualepillitteri (updated November 4, 2026), ArtificialWatch.