senn-techsenn-tech
KI
KI2026-10-24· By Franz Senn

Kimi K3.1 Pro: 4th in the AA Index, open weights announced, 2.8 trillion parameters

On 23 October 2026 Moonshot AI launched Kimi K3.1 Pro. The independent analyst Artificial Analysis (AA) scores it 55 points, 4th place in its Intelligence Index. Ahead of it are only Claude Opus 5.5 with 58, and Claude Sonnet 5.5 and Muse Wonder Full with 56 each. That is 11 points more than the predecessor K3. Moonshot also announced that the weights will follow by 4 November. For anyone who runs models in-house, the open question is what it takes to host this one.

Nvidia logo
The yardstick of this post is our AI server with four Nvidia RTX 5090 cards. The memory figures are in the section on the arithmetic. (Quelle: Simple Icons, file nvidia.svg (CC0))

We keep three kinds of statements apart. Verified means found in at least two independently read sources or retrieved directly. Vendor-claimed means it comes from Moonshot and has not been checked independently. Inferred is our own arithmetic.

What is measured: rank 4 and the price

The score of 55 and the ranking come from Artificial Analysis. The Decoder repeats them, so there is a single party measuring. Size and prices are verified: 2.8 trillion parameters in total, 104 billion of them active, 1 million tokens of context, text and images as input. Moonshot itself only talks about a "3T-class" model, according to Unite.ai. The figure of 2.8 trillion is not on any Moonshot page we read.

Artificial Analysis Intelligence Index, as of 23 October 2026Claude Opus 5.558Claude Sonnet 5.556Muse Wonder Full56Kimi K3.1 Pro55Kimi K344 · 55 minus 11060
Kimi K3.1 Pro takes 4th place with 55 points. The K3 value follows from the 11-point gap that Artificial Analysis states. (Quelle: Artificial Analysis, Kimi K3.1 Pro launch benchmarks)

The API costs 2 US dollars per million input tokens, 10 dollars per million output tokens and 0.20 dollars for cache hits. K3 was 3, 15 and 0.30 dollars, so the list price is 33 percent lower. AA puts one task of its index at 1.98 dollars, and K3 is about the same. The reason is that K3.1 Pro writes 55,000 output tokens per task instead of 48,000, which is 14 percent more. The lower token price is largely eaten up by the higher token count.

AA also measures the hallucination rate in its Omniscience test. It falls from 53 to 44 percent, and the Omniscience Index rises from 20 to 30 points. That is better, but 44 percent is still a high value.

Memory needed for Kimi K3.1 Pro (assumed same as K3) against the GPU memory of a four-card server, in GBVRAM 4x RTX 5090137 · measuredComputed 1 bit350 · inferredComputed 2 bit700 · inferredMeasured K3 (MXFP4)1561 · file size on Hugging FaceComputed 8 bit2800 · inferred02800
Only the VRAM and K3 bars are measured. The other three follow from the assumption of 2.8 trillion parameters and are our own arithmetic, because the K3.1 Pro weights do not exist yet. (Quelle: Hugging Face API moonshotai/Kimi-K3, nvidia-smi, own calculation)

What only the vendor says

The official launch video from Moonshot AI, 23 October 2026, 94 seconds. It is vendor marketing, so figures from it count as vendor-claimed.

All other benchmarks are vendor-claimed. The Next Web states explicitly that the figures have not been independently checked.

  • Terminal-Bench 4.0: Moonshot gives 65.7 percent, Opus 5.5 is at 66.4 and GPT-6 Astra at 58.2. AA measures 67 percent for K3.1 Pro. A different measurement, and no real contradiction.
  • Hallucinations: in its own Kimi Hallu Bench Hard, the rate falls from 23.4 to 8.3 percent according to Moonshot. Moonshot's blog itself says the values are no measure of the overall error rate in daily use.
  • Cyber capabilities: according to The Next Web, Moonshot says material on cyberattacks was removed from the training data. An internal exploit test gives 81 percent instead of 90.5 percent without the removal (K3: 33.3 percent). The model stays behind the best closed models. For a different test (ExploitBench), The Decoder lists 74.7 percent against 91.0 for Opus.
  • High-speed variant: Moonshot promises two to three times the speed. The Decoder gives 4 and 20 dollars for it.

The weak spot is easy to name. Apart from AA nobody has measured independently, and AA is a single measuring party. If you want a look for yourself, the Arena first-impressions video is a start.

First impressions from the Arena channel, 23 October 2026, about 42 minutes. It is not a measurement, and we take no figures from it.

The open-weights announcement

Unite.ai quotes Moonshot with "by November 4, 2026", and The Next Web and The Decoder report the same. That is vendor-claimed, and no weights exist yet: a search through the Hugging Face API on 24 October finds no moonshotai repo for K3.1, only a repo from a third party.

The track record gives one hint. According to New York University Shanghai (RITS), the K3 weights arrived on 27 July, the day Moonshot had promised. That makes the new announcement more credible, but does not prove it (inferred). The licence is also open. K3 ships under its own "Kimi K3 License": download, self-hosting, fine-tuning and quantization are allowed, and anyone reselling inference with more than 20 million US dollars in revenue over twelve months has to negotiate a separate agreement, according to NYU. For K3.1 Pro no source we read names the licence.

Why it will not run on our cards

The arithmetic has two parts, one measured and one derived.

The predecessor is measured: the repo moonshotai/Kimi-K3 lists 2,779,931,837,184 parameters and files totalling 1.561 terabytes. That is 0.56 bytes per parameter, just under 4.5 bits, which matches the native MXFP4 quantization that NYU mentions. If K3.1 Pro has the same size and architecture (an assumption, the weights are missing), the requirement is around 1.4 to 1.6 terabytes.

Against that stands a read-only query from 24 October: the AI server with four RTX 5090 cards reports 32,607 MiB four times via nvidia-smi, together 130,428 MiB, or about 137 GB. free shows 93,995 MiB of system RAM there, about 99 GB. Together that is roughly 235 GB, and most of the RAM is already taken by running services.

It follows (inferred) that 1.561 terabytes is 11.4 times the GPU memory. Even with all system RAM as overflow, a factor of about 6.6 is still missing. All experts have to be available because the selection changes per token. The 104 billion active parameters would be only about 52 GB at 4 bit, but that helps with compute and not with loading. To fit everything into 137 GB, each parameter could take 0.39 bits on average, before any KV cache. There is no usable format for that, and even 1 bit per parameter (0.35 terabytes by calculation) does not fit. For K3, NYU names a cluster of 64 or more accelerators as the realistic way to run it.

As a rule of thumb for the download check, parameters times bits divided by 8 gives the weights in bytes. What fits into 128 GB at 4 bit is roughly under 200 billion parameters (inferred, before KV cache and activations).

What admins can do now

From repo to decisionCheck the repofrom 4 NovemberRead the file sizesafetensors.totalParameters times bitsover 8weights in bytesCompare with VRAM137 GB hereDoes not fittest an API lane
The order of our check before a terabyte is downloaded. The licence is read in the first step. (Quelle: senn-tech, arithmetic of this post)
  1. From 4 November, check the repo. Before anything is downloaded, read the file size (safetensors.total in the Hugging Face API) and the licence text. The licence of K3.1 Pro is unknown.
  2. Test it as an API lane, not locally. Thinking mode is mandatory, and without it the API answers with HTTP 400, according to Unite.ai. The 14 percent extra output tokens belong in the cost estimate. Business or personal data does not go to a third-party host without review.
  3. Calculate first, download second. The arithmetic above is the template before a terabyte goes through the line.

One side note for context: The Next Web mentions the White House allegation from July that Moonshot used banned Nvidia chips and copied an Anthropic model. The report names no evidence. Moonshot has not made this statement; it is an allegation by others.

Further sources

Questions?
Are the Kimi K3.1 Pro weights released yet?+

No, not as of 24 October 2026. According to Unite.ai, The Next Web and The Decoder, Moonshot AI announced it will release the weights by 4 November 2026. That is a vendor claim. Artificial Analysis only says the weights are expected soon after launch. A Hugging Face search on 24 October shows no moonshotai repo for K3.1, only a third-party repo.

How big is Kimi K3.1 Pro, and can you run it locally?+

Artificial Analysis and The Decoder give 2.8 trillion parameters, 104 billion of them active. The predecessor K3 takes 1.56 terabytes on Hugging Face. If K3.1 Pro is the same size (an assumption), loading it alone needs more than eleven times the GPU memory of a server with four RTX 5090 cards. That is unrealistic for an office homelab, but an API connection works.

What does Kimi K3.1 Pro cost through the API?+

According to Artificial Analysis and The Decoder, 2 US dollars per million input tokens, 10 dollars per million output tokens and 0.20 dollars for cache hits. That is 33 percent less than K3. Thinking mode is mandatory according to Unite.ai, and the model produces 14 percent more output tokens per task than K3, so the saving is smaller than the list price suggests.