Moonshot's Kimi K3.1 Pro Is a 3T-Class Open-Weight Model

On October 23, 2026, Beijing-based Moonshot AI introduced Kimi K3.1 Pro, the first model in its K3.1 series and, in the company's words, its most capable 3T-class open-weight model to date. In its technical blog, Moonshot describes K3.1 Pro as a post-trained upgrade to Kimi K3 with substantially larger post-training compute, including large-scale joint reinforcement learning across hundreds of downstream domains. The model ships with a 1-million-token context window (the Kimi API lists 1,048,576 tokens) and targets coding, 3D creation, financial research, and long-horizon agent work. A faster sibling, Kimi K3.1 Pro Highspeed, offers the same quality at roughly 2–3× the speed, per Kimi Code release notes. Caixin Global covered the release on October 26, calling it Moonshot's most powerful model yet as domestic developers try to keep pace with a rapid series of upgrades from U.S. rivals; the outlet reports Moonshot said the model has about 3 trillion parameters—a figure the official tech blog does not spell out, instead calling it 3T-class. Moonshot's Kimi Code release notes put the base Kimi K3 at 2.8 trillion parameters.

Key facts

What Moonshot claims on benchmarks

The figures below come from Moonshot's K3.1 Pro tech blog and should be read as company-reported; footnotes there describe harnesses, judge models, and leaderboard sourcing.

Moonshot states that across coding, agentic, and knowledge suites, K3.1 Pro matches or approaches the strongest proprietary models on the benchmarks it highlights, while noting that proprietary testing may involve fallback behavior on some internal coding benches.

Post-training focus: coding, 3D, and finance

Moonshot attributes K3.1 Pro's gains to stack-wide RL investments—inference, sandboxes, rollout scheduling, and environment QA—so additional RL compute returns predictable improvements. The blog emphasizes end-to-end coding (game engines such as Unity, Unreal, and Godot; long-horizon debugging), 3D modeling via scripts for Blender and CAD tools, and financial workflows spanning modeling, research, and simulated business operations. On internal finance benches, Moonshot reports K3.1 Pro improving on K3 by 7.2 points on Financial Modeling, 11.6 on Financial Research, and reaching about 4.2× K3's score on Business Operation. For factual stress-testing, Moonshot cites Kimi Hallu Bench Hard: claim-level hallucination rate falling from 23.4% on K3 to 8.3% on K3.1 Pro, with responses containing at least one factual error at 42.4% of K3's rate—rates Moonshot stresses do not represent everyday error rates.

Open weights, cyber capability, and release policy

Moonshot plans to publish full model weights by November 4, 2026. Ahead of that open release, the company says it identified and removed cyber-related data from both mid-training and RL stages while keeping general coding tasks. On its in-house xdev-bench-hard suite, Moonshot reports the unfiltered training pipeline would reach 90.5% pass@3 versus 81.0% for the shipping K3.1 Pro—evidence, in Moonshot's view, that the controls lower offensive cyber skill relative to what the same pipeline could produce. Moonshot's open-weight principles state it aims not to release open weights that extend the absolute frontier of cyber capability, that K3.1 Pro's offensive cyber ability is by design below what its pipeline can produce, and that it remains behind frontier proprietary models that have been public for at least three months, while intending to progressively raise cyber capability in future open releases.

Availability and API pricing

K3.1 Pro and Highspeed are live on Kimi.ai, Kimi mobile apps, Kimi Work (version 3.1.0+ on Windows and Apple silicon Macs), Kimi Code (select via /model), and the Kimi API with model IDs kimi-k3.1-pro and kimi-k3.1-pro-highspeed. Kimi Enterprise admins can enable the models, but access is disabled by default at launch. API customers can use Zero Data Retention, per the tech blog.

Per official pricing, both variants use flat per-token rates across the entire context window—no length-based tiers:

The pricing page notes standard K3.1 Pro undercuts current Kimi K3 API list prices ($3.00 / $15.00 / $0.30 per million for input, output, and cached input). Thinking cannot be disabled; lowering reasoning_effort may reduce token use on simpler tasks.

What it means for developers

For teams building agents or coding assistants, K3.1 Pro sits in an unusual spot: frontier-leaning benchmark claims and a million-token window at API prices, with open weights arriving within days—but only after Moonshot deliberately caps cyber skill relative to its internal maximum. That trade-off matters for security researchers who want auditable models on their own infra, and for red teams who should not assume the public checkpoint equals Moonshot's unconstrained pipeline.

Self-hosting a 3T-class model remains realistic mainly for large cloud budgets or research clusters; most developers will integrate through the Kimi API or Moonshot's apps, using Highspeed when latency dominates and the standard tier when cost does. Until weights land, treat capability and safety claims as tied to Moonshot's hosted endpoints; after Nov. 4, compare the released artifact to the API behavior you measured in production.

Sources

• Moonshot AI: Introducing Kimi K3.1 Pro — Oct. 23, 2026
• Kimi API Platform: Kimi K3.1 Pro — Oct. 23, 2026
• Moonshot AI: Kimi K3.1 Pro Pricing Explained: Membership Plans and API Costs — Oct. 24, 2026
• Kimi Code Docs: What's New — Oct. 23, 2026
• Caixin Global: Moonshot AI Launches More Powerful Kimi Model — Oct. 26, 2026