On this page
Kimi K3.1 Pro is Kimi K3 with more training on top, not a bigger model. Its maker, Moonshot AI, put the downloadable weights on Hugging Face this week, and TechNode reported the release on 5 November 2026. It keeps the size and licence terms of K3. It costs a third less on Moonshot's API, and it scores higher on every coding test the two models ran the same way.
If you call K3 through Moonshot's API, switch: the new model is cheaper and stronger.
If you run K3 on your own GPUs, plan a redeploy. The download is the same size, but one layer type changed.
If you need prompts processed in Canada, either model works, because you can download and run both yourself.
Key facts
| Kimi K3.1 Pro | Kimi K3 | |
|---|---|---|
| What it is | K3 after extra reinforcement learning | Moonshot's base flagship |
| Size | 2.8T parameters, 104B used per token | 2.8T parameters, 104B used per token |
| How much text it reads at once | 1,048,576 tokens | 1,048,576 tokens |
| Weights on Hugging Face | November 2026 (first commit 2 Nov) | July 2026 |
| Download | 96 files, 1.56 TB | 96 files, 1.56 TB |
| Licence | Kimi K3.1-Pro License | Kimi K3 License |
| Moonshot API price, per million tokens | $2 in, $10 out | $3 in, $15 out |
A token is about three quarters of a word. Open weights means you can download the model files and run them on your own hardware.
What changed this week?
The weights landed. The launch post put the new model in Kimi's apps and API on 23 October. It promised open weights by 4 November.
The Hugging Face repository's first commit is dated 2 November, per its commit history, and a small fix followed on 5 November. TechNode reported the release on 5 November.
If you self-host, the upgrade is no longer API-only: you can now run it yourself. Our July comparison of K3 and K2.7 Code is now one version behind.
What changed from Kimi K3 to Kimi K3.1 Pro?
The training changed, and so did one type of layer. The size did not. The model card says the new model was post-trained from K3 with "large-scale joint reinforcement learning across hundreds of downstream domains." In plain terms, it practised many more tasks and was scored on each result.
Three differences are concrete: attention, training data and a faster variant.
Attention. The model has 93 layers. In the original, 24 of them read the whole context at every step. The new model swaps those 24 for a sparse design from DeepSeek that reads only part of the context. Moonshot says the aim is cheaper work on long inputs. The K3 card shows the other layers are unchanged.
Cyber training data. The launch post says cyber-related data was taken out of training for the open-weight release.
A faster variant. A Highspeed version, called kimi-k3.1-pro-highspeed on the API, runs 2 to 3 times faster, per the quickstart.
Everything else carries over, including image input and the compact 4-bit weight format.
Did the licence change?
Only the name you would have to display. On 5 November we compared the new licence with the old one line by line, and the conditions match.
The K3 and K3.1 Pro licences are custom Moonshot licences. Each opens with wording borrowed from the MIT open-source licence, which is why some coverage calls them "modified MIT". Then each adds two conditions:
| Condition | Who it applies to | What it asks |
|---|---|---|
| Model as a Service | Companies selling API or fine-tuning access, with more than $20M revenue over any 12 months | Sign a separate agreement with Moonshot before commercial use |
| Display | Products with more than 100M monthly active users, or more than $20M monthly revenue | Show the model's name prominently in the interface |
Internal use is exempt from both. The licence defines internal use as anything that does not make the model, its outputs or its capabilities available to third parties. A company running the model on its own GPUs for its own staff is covered by that exemption. If you plan to resell access to the model, read the full text first.
What do the coding numbers say?
The new model scores higher on every coding test where both ran the same version. Every number here is the vendor's own, from the two model cards and the launch post. Nobody independent has rerun them yet.
| Task, in plain terms | Kimi K3.1 Pro | Kimi K3 | Same setup? |
|---|---|---|---|
| Moonshot's in-house coding suite, v2.0, run inside Claude Code | 77.1 | 73.7 | Yes |
| Moonshot's long-horizon coding suite, v3.0 | 61.2 | 55.95 | Vendor's own side-by-side |
| Long, multi-file software tasks, v1.1 task set | 75.5 (Claude Code) | 67.5 (Kimi Code) | Same tasks, different tool |
The first row is the cleanest comparison. Same suite, same version, same coding tool, and a 3.4-point gain. The third row shares a task set, but each model ran in a different coding tool. Part of that 8-point gap may come from the tool.
Several headline numbers do not line up, so we left them out of the table.
Terminal tasks. K3 reports 88.3 on Terminal-Bench 2.1. The new model reports 65.7 on Terminal-Bench 4.0. Those are different tests, so the lower number is not a step back. Our post on Terminal-Bench version drift explains why.
ProgramBench. The two cards report very different scores. The newer card counts a task as solved only above a 95% threshold. The older card does not say it used that rule.
SWE-Marathon and GDPval. Each model ran a different version of these two tests.
Outside coding, the launch post reports a drop in made-up claims on its hardest in-house test, from 23.4% to 8.3%.
How much does Kimi K3.1 Pro cost on Moonshot's API compared with K3?
The new model costs $2 per million input tokens and $10 per million output tokens. K3 costs $3 and $15. The vendor's pricing page loads its table only in a browser. So we took both rates from Moonshot's own listings on OpenRouter, as archived by the Wayback Machine: K3.1 Pro as of 4 November 2026 and K3 as of 19 October 2026.
| Job | Kimi K3.1 Pro | Kimi K3 |
|---|---|---|
| One long agent run: 1M tokens in, 200K out | $4.00 | $6.00 |
| A month of team coding help: 500M in, 50M out | $1,500 | $2,250 |
| Same month through the Batch API (60% of list) | $900 | $1,350 |
Moonshot's platform changelog added both models to the Batch API in October, at 60% of standard rates. The pricing page, as of 23 October, says repeated input is cached and billed at a lower rate automatically. That cuts the bill further for agents that resend the same files.
We could not find a Highspeed rate on a Moonshot page, so we left it out. The new model opens on the API after a $1 minimum top-up, per the quickstart as of 4 November.
Should you switch from Kimi K3 to Kimi K3.1 Pro?
On Moonshot's API, yes. It is cheaper and scores higher, and switching means changing one model field. Self-hosting takes more work.
| You run K3 | Move to K3.1 Pro? | What to check first |
|---|---|---|
| Through Moonshot's API | Yes | Rerun your own prompts, since retraining changes behaviour |
| Through another host | When your host lists it | On 4 November, OpenRouter hosts charged $1.80 to $2.30 per million input tokens |
| On your own GPUs | Yes, as a planned redeploy | Your serving engine supports the new sparse-attention layers |
| For security research | Test first | Cyber data was removed from the open-weight training run |
We have not run the new model on our own tasks yet, so the coding numbers above are the vendor's alone. Retraining also changes behaviour even when scores rise, so prompts tuned for K3 may need adjusting.
For the field before this release, see our August roundup of open-source coding models and how K3 compared with GLM-5.2. For the other open-weight release of the past fortnight, see MiniMax M3.1 Flash vs DeepSeek V4.1 Flash.
Can you run Kimi K3.1 Pro in Canada?
Yes. Both models are open-weight, so either can run on dedicated GPUs in a Canadian cloud region or on hardware you own. On AWS, that means ca-central-1 in Montreal or ca-west-1 in Calgary.
Three facts set the size of the job.
Download. Each model is 96 weight files totalling 1.56 terabytes, per the Hugging Face listing as of 5 November.
Licence. Internal use is exempt from both licence conditions. Reselling access past $20M in yearly revenue needs an agreement with Moonshot.
Hardware. The model card gives no GPU guidance. It names vLLM, SGLang and TokenSpeed as serving engines. Your workload sets the size of the deployment.
Our Canadian residency guide lists the trade-offs. Large GPU instances are scarcer in Canadian regions than in American ones. You pay for GPUs while they sit idle. And you own the serving, patching and safety filtering.
XY Space deploys open-weight models on dedicated GPUs in the client's own Canadian cloud account, or on-premises, so prompts and outputs stay in the country. Then we build the workflow on top. Your people run the work and the model takes part. See how we deploy, and our view on running open models at frontier level.
If you need this model's capability with inference kept in Canada, bring one job to a free 30-minute call.
FAQ
Is Kimi K3.1 Pro better than Kimi K3 for coding?
On the vendor's own numbers, yes. In Moonshot's in-house coding suite, run inside Claude Code, the new model scored 77.1 and K3 scored 73.7. It also led on a long-horizon suite and on a shared set of long, multi-file tasks, though for that last one the two ran in different coding tools. Nobody independent has rerun these yet, so test on your own code first.
Is Kimi K3.1 Pro the same size as Kimi K3?
Yes. Both have 2.8 trillion parameters and use 104 billion of them per token. Both read up to about a million tokens at once and download as 1.56 terabytes. The new model is K3 with more reinforcement learning. One part inside changed: 24 layers that read the whole context now read only part of it, which Moonshot says makes long inputs cheaper.
Is Kimi K3.1 Pro open source under MIT?
No. It ships under the Kimi K3.1-Pro License, a custom licence that starts with MIT-style wording and adds two conditions. Companies that sell model access and earn more than $20M a year need a separate agreement. Very large products must display the model's name. Internal use is exempt from both conditions, and the terms match the K3 licence.
Why is Kimi K3.1 Pro's Terminal-Bench score lower than Kimi K3's?
Because the two ran different versions of the test. K3 reported 88.3 on Terminal-Bench 2.1, and the new model reported 65.7 on Terminal-Bench 4.0. Scores from different versions cannot be compared, so the gap says nothing about the new model being weaker. Moonshot has not published either model's score on the other version.
How much cheaper is Kimi K3.1 Pro than Kimi K3?
About a third. On Moonshot's API, the new model costs $2 per million input tokens and $10 per million output tokens. K3 costs $3 and $15. Those rates come from Moonshot's listings on OpenRouter, archived on 4 November 2026 for the new model and 19 October 2026 for K3. A month of 500 million input and 50 million output tokens costs $1,500 on the new model and $2,250 on K3.
Sources
- Moonshot AI, Kimi K3.1 Pro model card, Hugging Face, as of the 5 Nov 2026 commit (history); card, licence and weights unchanged since the 2 Nov first commit. Architecture, MXFP4 quantization, vendor scores: Terminal-Bench 4.0, DeepSWE v1.1, ProgramBench, Kimi Code Bench 2.0, SWE-Marathon v1.1, GDPval-AA v2.1.
- Moonshot AI, Kimi K3.1 Pro License, as of the 5 Nov 2026 commit.
- Moonshot AI, Kimi K3 model card, Hugging Face, as of the 2 Sep 2026 commit (history). Terminal-Bench 2.1, DeepSWE v1.1 tasks, Kimi Code Bench 2.0, SWE-Marathon pre-release branch, GDPval-AA v2.
- Moonshot AI, Kimi K3 License, unchanged since the 27 Jul 2026 first commit.
- Moonshot AI, Kimi K3.1 Pro launch post, 23 Oct 2026, last modified 24 Oct. Kimi Code Bench 3.0 Long-horizon, Kimi Hallu Bench Hard, open-weight date.
- Moonshot AI, Kimi K3.1 Pro quickstart, Wayback snapshot of 4 Nov 2026; pricing page, snapshot of 23 Oct 2026; platform changelog, October 2026 entry.
- OpenRouter, Kimi K3.1 Pro providers, Wayback snapshot of 4 Nov 2026, and Kimi K3 providers, snapshot of 19 Oct 2026.
- TechNode, Moonshot AI releases Kimi K3.1 Pro model weights, 5 Nov 2026.
Written by
Cho Yin Yong
Principal AI Solutions Engineer, XY Space
Principal AI Solutions Engineer at XY Space. University of Toronto lecturer for five years, co-author of two patents, winner of two competitive AI awards, and nine years of regulated engineering leadership.
More from Cho Yin YongShare this article
Work with us
We build the systems these posts describe, and we'll tell you in the first call whether yours is worth building.
Start a project