Kimi K3.1 Pro: 4th on the Intelligence Index, But No Weights Until Nov 4

The short version

On October 23, 2026, Beijing-based Moonshot AI released Kimi K3.1 Pro, a post-trained successor to Kimi K3. Independent evaluator Artificial Analysis scored it 55 on the Intelligence Index — fourth of 215 models, behind only Claude Opus 5.5 (58), Claude Sonnet 5.5 (56), and Meta's Muse Wonder Full (56). That is an 11-point jump over K3.

Here is the part that matters more than the ranking: the weights are not out yet. Moonshot says full weights will be released "by Nov 4, 2026." As of this writing there is no K3.1 repository in Moonshot's Hugging Face organization and no published license. Until that day, "open-weight" describes an intention, not an artifact.

This piece separates what is measured today from what changes on November 4.

Specs and pricing: same size, one-third cheaper

2.8T total parameters, 104B active per token (MoE: 896 experts, 16 selected plus 2 shared)
1,048,576-token context, text and image input, text output
93 layers, of which 69 are linear-attention KDA layers and 24 are gated MLA
Quantization-aware training: MXFP4 weights / MXFP8 activations
API: $2 per million input tokens, $10 output, $0.20 cached input — a 33% cut from Kimi K3 ($3/$15/$0.30)
K3.1 Pro Highspeed costs $4/$20 and Moonshot claims 2–3x the throughput

K3.1 Pro is treated as a flagship: the API unlocks only after a top-up of at least $1. On Kimi.ai and Kimi Work the free tier gets "limited usage," and Highspeed requires the Pro plan ($39/month, or $31/month billed annually). Consumer tiers run Plus $19, Pro $39, Max $99, Ultra $199 monthly.

What independent testing actually shows

The Artificial Analysis numbers are more useful than the vendor chart, because they expose where the gains are concentrated.

GDPval-AA: 1,846 Elo (up 312 from K3), third overall behind Opus 5.5 (1,866) and Muse Wonder Full (1,857)
Terminal-Bench 4.0: 67% — up from 13% for K3, ahead of Sonnet 5.5 (64%) and, on AA's measurement, the best result recorded on this benchmark. Moonshot's own figure is 65.7%, versus Opus 5.5 at 66.4% and GPT-6 Astra at 58.2%
AutomationBench-AA: 58% → 73% on agentic SaaS workflows
Weaker on knowledge evals: Humanity's Last Exam 47% → 52%, SciCode 59% → 61%, CritPt 23% → 25%
Hallucination rate: 53% → 44% on AA-Omniscience (accuracy 48% → 51%)

Read the direction, not the ranking. This is a model that got dramatically better at acting in software and terminal environments, and only modestly better at knowing things. If your workload is agentic coding, the jump is real. If it is open-ended research, the improvement is incremental.

The cost surprise: 33% cheaper tokens, the same task price

AA measured 55,000 output tokens per Intelligence Index task, 14% more than K3's 48,000. Because the model writes more, the effective cost per task is $1.98 versus $2.00 — a rounding error, not a saving. K3.1 Pro is also slow at 31 tokens per second (141st of 186 models measured, against a median around 82.5 TPS).

That profile is fine for batch agent work and painful for interactive chat. Budget by task, not by price sheet.

November 4: three things to check before you commit

1. The license. Kimi K3 shipped under a custom "Kimi K3 License," not MIT or Apache. It requires a separate agreement with Moonshot for Model-as-a-Service operators whose group revenue exceeds $20M in any consecutive 12 months, and UI attribution ("Kimi K3" displayed prominently) for products above 100M MAU or $20M in monthly revenue. Internal use and access through Moonshot's own products or certified partners are exempt. Whether K3.1 Pro inherits the same terms is unknown — no license file exists yet. 2. The hardware floor. The K3 MXFP4 checkpoint sums to 1,560.9 GB across 96 safetensors shards. Eight H200s (1,128 GB aggregate) are 433 GB short before a single KV cache block; eight B200s (1,440 GB) still do not fit. The practical floors are 8x B300-class GPUs at 288 GB or two nodes of 8x H200. vLLM's recipe asks for at least 8x GB300, and 32 H100s if you insist on Hopper. Add KV cache (13.5 GiB per sequence at FP8 for a full 1M-token context) and a flat 213–427 MiB of recurrent state per concurrent request for the 69 KDA layers. Long context is cheap here; concurrency is what costs money. 3. The cyber capability they removed on purpose. Moonshot states it stripped cyber-related data from both mid-training and the RL environment mix. On its internal xdev-bench-hard, the final K3.1 Pro solves 81.0% of tasks; an ablation checkpoint trained without the redaction solves 90.5%. For context, Kimi K3 scored 33.3% and DeepSeek V4.1 Flash reaches 85.7%. The company's stated position is that no open release should extend the absolute frontier of offensive cyber capability. Once the weights exist, researchers can test how durable that 9.5-point gap is under fine-tuning. A third-party Hugging Face page already advertises an abliterated K3.1 build for release day — which removes refusal behavior but cannot restore capabilities never trained in.

Why the 3D work is the underrated part

K3.1 Pro was explicitly trained to drive 3D tools through Python scripts, so Blender, CAD packages, Unity, Unreal Engine, and Godot work without MCP connectors. Moonshot's examples include building a full Godot scene from scratch and assembling architectural models in Blender. A new 3D Make mode in Kimi.ai and Kimi Work previews, edits, and exports models for 3D printing, video, or Live Photos.

The division of labor is sensible: parametric modeling (mechanical parts, furniture, buildings) goes through code the model writes; mesh generation for characters goes to diffusion tools like Tripo, which Kimi calls automatically with prepared prompts and reference images. For small studios, that is a materially different workflow from "generate an image and hope."

What to actually do this week

Try the API now if you have agentic or batch workloads that tolerate 31 TPS. The $1 top-up is a low-cost test, and Terminal-Bench 67% is the number to validate against your own tasks.
Wait for November 4 before any production commitment. The license, the real checkpoint size, and whether a non-quantized build ships all land that day.
Do not plan self-hosting. A 1.5 TB checkpoint in 8-GPU increments is a cloud-scale decision, not a laptop one.
Treat 44% hallucination as a real number. It is a 9-point improvement, not a solution. Keep retrieval and citation checks in the loop for finance and legal work.

Bottom line

Kimi K3.1 Pro may be the moment open weights closed the gap to 3 points on the Intelligence Index — a claim AA supports with its own testing, not just Moonshot's. But the honest state of play on November 1 is this: you can use a very strong model through an API today, and you cannot download one yet. The interesting event this week is not the launch. It is what shows up in that Hugging Face organization on November 4.