What Is Kimi K3.1 Pro? Features, Benchmarks and API

Kimi K3.1 officially launched on October 23, 2026, as Kimi K3.1 Pro, resolving weeks of speculation that began with a cryptic Pi-digit teaser. Developed by Moonshot AI, the model is a post-trained 2.8-trillion-parameter sparse mixture-of-experts (MoE) system with 104 billion active parameters and a 1-million-token context window. Released alongside a faster K3.1 Pro Highspeed variant, the model is live today across Kimi web, mobile, Kimi Code, and API endpoints, and its open weights are now published on Hugging Face under the Kimi K3.1 Pro License. Kimi K3.1 Pro is now live on Kie.ai, where you can call it through the Kimi K3.1 Pro API.

Updated 2026-10-21: Kimi K3.1 support appeared in a briefly opened SGLang pull request, whose test data points to a new "medium" thinking tier and a DeepInfra deployment trace dated October 18 (see the Update below).

Key Takeaways

• Moonshot AI officially released Kimi K3.1 Pro and K3.1 Pro Highspeed on October 23, 2026.
• The model keeps K3's scale at 2.8 trillion total parameters (104B active parameters per token) and a 1-million-token context window, enhanced by large-scale joint reinforcement learning.
• On the Artificial Analysis Intelligence Index, K3.1 Pro debuted in 4th place worldwide with a score of 55, landing 11 points above Kimi K3 and just behind Claude Opus 5.5 (58).
• Key benchmark results include 65.7% on Terminal-Bench 4.0, 92.7% on BenchCAD with tools, and a sharp reduction in claim-level hallucinations from 23.4% to 8.3%.
• Official API pricing dropped by 33% relative to K3: $2.00 per 1M input tokens and $10.00 per 1M output tokens ($0.20 per 1M cached tokens).
• Open weights are published on Hugging Face (moonshotai/Kimi-K3.1-Pro) under the Kimi K3.1 Pro License.

What Is Kimi K3.1?

Kimi K3.1 is the flagship post-trained successor to Kimi K3, officially released by Moonshot AI on October 23, 2026, under the name Kimi K3.1 Pro. While the "3.1" label initially signaled an incremental update, Moonshot's release delivers substantial improvements in long-horizon coding, 3D procedural generation, and autonomous agent workflows.

The earliest confirmed vendor signal was a Pi reference. On September 19, 2026, Moonshot’s official account published a mysterious number string that community members decoded as the decimal expansion of π with the opening “3.1” stripped out. Ananth’s report on the teaser and JAZII’s decoding of the string provided the first concrete clues that K3.1 was in active development.

Source: @Ananth7e

Earlier user feedback—such as a July 18, 2026 comment attributed to Moonshot researcher Xinyu Yang—highlighted user demands for faster inference, more token-efficient reasoning on large tasks, and higher reliability. The final release addresses those demands directly through extensive joint RL rather than architectural parameter expansion.

For background, the original Kimi K3 established Moonshot's 2.8-trillion-parameter sparse mixture-of-experts foundation with a 1,048,576-token context window. The earlier Kimi K3 overview covers that baseline in detail. K3.1 Pro retains that physical parameter footprint while overhauling the post-training stack.

Kimi K3.1 at a Glance

Specification	Current status
Developer	Moonshot AI
Type	Sparse Mixture-of-Experts (MoE) reasoning model (2.8T total, 104B active parameters)
Modality	Multimodal (text, image input)
Context window	1,048,576 tokens (1M context)
Pricing	Pro: $2.00 / 1M input, $10.00 / 1M output ($0.20 cache); Pro Highspeed: $4.00 / $20.00
Availability	Live on Kimi.ai, Kimi Work, Kimi Code, and Kimi API; open weights on Hugging Face
License	Kimi K3.1 Pro License (open weights on Hugging Face)
Expected position	Successor to Kimi K3; top-ranking open-weights model on Intelligence Index
Release date	October 23, 2026

How Kimi K3.1 Works / What Makes It Different

Kimi K3.1 Pro is built as a heavily post-trained evolution of Kimi K3, utilizing substantially increased post-training compute. Moonshot optimized its reinforcement learning framework, long-context inference pipeline, sandbox infrastructure, and rollout scheduling to enable large-scale joint RL across hundreds of downstream domains.

Architecturally, K3.1 Pro retains K3’s core foundations:

• Stable LatentMoE Routing: Routes tokens to 16 active experts and 2 shared experts out of 896 total experts, operating at 104 billion active parameters per token.
• Kimi Delta Attention (KDA) & Gated MLA: Combines 69 KDA linear-complexity layers with 24 Gated Multi-Head Latent Attention layers for high-precision long-context recall across 1 million tokens.
• Attention Residuals: Selectively retrieves representations across depth layers to maintain training and inference efficiency.

Operationally, K3.1 Pro introduces several critical changes confirmed by testing and developer traces:

1. Granular Thinking Effort: As first revealed in SGLang PR test data, K3.1 adds a "medium" thinking-effort tier alongside low, high, and max (with max remaining default).
2. K3.1 Pro Highspeed Variant: Ships alongside the base model, running 2× to 3× faster for interactive coding and conversational speed.
3. Pre-release Canary Fingerprint: Community testing under the anonymous Arena moniker "Kibitz" demonstrated byte-identical tokenizer and special-token behavior matching Kimi's proprietary stack.
4. Deliberate Offensive Cyber Pruning: Moonshot voluntarily stripped offensive exploit training data before shipping open weights, dropping its ExploitBench score from 90.5% to 74.7% (or 81.0% in independent exploit evaluations) to curb automated vulnerability weaponization.

What You Can Do With Kimi K3.1

With official endpoints live and open weights published, developers can integrate K3.1 Pro across several primary workloads:

• Long-Horizon Coding & Multi-Step Execution: K3.1 Pro handles multi-file refactoring, end-to-end debugging, and complex terminal tasks. It scores 65.7% on Terminal-Bench 4.0 (up from K3's 13%) and 61.2% on Kimi Code Bench 3.0 Long-horizon.
• Procedural 3D Modeling and Game Assets: A major focus of the release is spatial reasoning and 3D generation. K3.1 Pro generates functional Three.js scenes, CAD scripts, and game engine assets (Unity, Unreal Engine, Godot), scoring 92.7% on BenchCAD with tools.
• Autonomous Agent Workflows: Designed for long execution chains, the model reached #11 on Agent Arena (+7.6% net jump over K3) with an average task cost of $1.03.
• Financial Research and Enterprise QA: Claim-level hallucinations dropped from 23.4% on K3 down to 8.3% on Kimi Hallu Bench Hard, yielding gains in financial modeling (+7.2 points) and leading OfficeQA Pro v2 (64.4%).
• Structured Output and Tool Calling: Native support for tool use, multi-modal image input, and strict JSON formatting integrated directly into Kimi Code and API SDKs.

How Kimi K3.1 Compares

Kimi K3.1 Pro moves Moonshot into direct competition with frontier proprietary models like Claude Opus 5.5 and GPT-6 Astra, while dramatically undercutting them on price per token.

Benchmark / Metric	Kimi K3.1 Pro	Kimi K3	Claude Opus 5.5	GPT-6 Astra
Artificial Analysis Index	55 (#4)	44	58 (#1)	—
Terminal-Bench 4.0	65.7%	13.0%	66.4%	58.2%
BenchCAD (with tools)	92.7%	—	96.2%	95.9%
SWE Marathon v1.1	51.7%	—	46.3%	42.5%
GDPval-AA v2.1 (Elo)	1,846	1,534	1,866	1,573
Context Window	1,000,000	1,000,000	200,000+	1,000,000
API Price (Input/Output / 1M)	$2.00 / $10.00	$3.00 / $15.00	~$15 / ~$75	~$5+ / ~$15+

K3.1 Pro delivers roughly 66.7% of K3's API cost ($2 / $10 vs $3 / $15 per million tokens), matching or beating proprietary systems on coding benchmarks like SWE Marathon while trailing Claude Opus 5.5 by less than a point on Terminal-Bench 4.0. The Kimi K3 versus Claude analysis provides further context on how Moonshot's models compare against Anthropic's flagship lineup.

Availability: How to Access Kimi K3.1

Kimi K3.1 Pro is now broadly accessible across Moonshot's ecosystem, with third-party aggregators and downloadable weights rolling out:

Access route	Current status	Engineering implication
Moonshot Official App & Web	Live on Kimi.ai and Kimi Work	Available for consumer chat and workspace projects
Kimi Code	Live globally	Tailored for agentic coding and interactive 3D Three.js execution
Official Kimi API	Live (kimi-k3.1-pro, kimi-k3.1-pro-highspeed)	OpenAI-compatible endpoint at $2/$10 per 1M tokens ($0.20 cache)
Public Model Weights	Published on Hugging Face	Full 2.8T MoE weights for self-hosted deployment, with vLLM, SGLang and TokenSpeed recommended by Moonshot AI
kie.ai API & Third Parties	Live via OpenRouter and Kiro; K3 available as baseline	Developers can integrate or benchmark via the Kimi K3 API page

What We Don't Know Yet

While the launch answered questions regarding model scale, pricing, and benchmarks, several practical items remain open:

• Subscription Rate Limits: Early user feedback on Kimi's consumer and developer platforms reported restrictive message quotas on paid plans, with heavy tasks quickly exhausting allocations.
• Commercial Influencer Scrutiny: Leaked marketing documents following the launch prompted discussions regarding disclosure standards for overseas promotional campaigns.
• The Rumored 5T–6T Model: Nomura reports and developer leaks referenced a distinct, massive 5-to-6-trillion parameter model also in development at Moonshot, whose compute constraints and timeline remain unannounced.

Frequently Asked Questions

What is Kimi K3.1?

Kimi K3.1 (officially launched as Kimi K3.1 Pro) is Moonshot AI's flagship 2.8-trillion-parameter sparse mixture-of-experts model, released on October 23, 2026. It features 104 billion active parameters, a 1-million-token context window, and major upgrades in coding, 3D generation, and long-horizon agentic workflows.

Is Kimi K3.1 released?

Yes. Kimi K3.1 Pro and K3.1 Pro Highspeed officially launched on October 23, 2026, and are live on Kimi.ai, Kimi Work, Kimi Code, and the Kimi API.

Is Kimi K3.1 open source?

Kimi K3.1 Pro is an open-weight model. Hosted access is live, and the full model weights are published on Hugging Face under the Kimi K3.1 Pro License, a Moonshot-specific license rather than a standard open-source license such as MIT or Apache 2.0.

How much does Kimi K3.1 cost?

Kimi K3.1 Pro API pricing is $2.00 per 1 million input tokens and $10.00 per 1 million output tokens ($0.20 per 1M cached input tokens), a 33% discount compared to Kimi K3. The K3.1 Pro Highspeed variant is priced at $4.00 input and $20.00 output per 1 million tokens.

What is the Kimi K3.1 context window?

Kimi K3.1 Pro features a confirmed 1-million-token (1,048,576 tokens) context window, maintaining full long-context parity with Kimi K3.

Kimi K3.1 vs Claude Fable 5: which is better?

On official benchmarks, Kimi K3.1 Pro ranks among top frontier models, scoring 55 on the Artificial Analysis Intelligence Index (behind Claude Opus 5.5 at 58 and Sonnet 5.5 at 56) and 65.7% on Terminal-Bench 4.0. While Claude models retain an edge on certain strict reasoning tasks, K3.1 Pro delivers comparable agentic and 3D output at roughly one-third the token cost.

Update — 2026-10-17

The most specific release claim so far comes from a Nomura Securities report, summarized on October 10. It said Kimi K3.1 would launch before mid-October. It also said the model would focus on agents that carry out long tasks over many steps, using more post-training, computer-operation trajectories, and real-world task data. A leaker disputed the timing the same day, saying K3.1 would not launch before October 15. Aggregator posts then moved the window to "mid-to-late October" (@Mr_Salio, @0x0SojalSec). As of October 17, no release had occurred, and Moonshot had made no official statement. The same reports describe a separate, much larger Moonshot model in development, constrained by a shortage of high-end training compute. Its size is reported inconsistently, as either 5–6 trillion or 3–6 trillion parameters. It should not be confused with K3.1.

Several single-source access claims also appeared:

• A K3.1 model ID reportedly appeared in Moonshot's API with a 1-million-token context preview.
• An alleged API configuration leak showed three reasoning-intensity tiers and multi-agent collaboration features.
• K3.1 was claimed to have been briefly listed on September 30 and then withdrawn.

None of these has been verified, and none comes with benchmarks.

One possible regulatory factor also surfaced. Draft Chinese export rules would require a commerce-ministry license before open-sourcing models with at least 2 trillion parameters or above a training-compute threshold. Existing models would be exempt, and the comment period runs until November 15. If the rules are adopted, they could affect an open-weight K3.1 release. However, they are still a draft and not in force, and no source has shown that they caused the current delay.

Update — 2026-10-21

The clearest technical trace so far came from the inference framework SGLang on October 20. A pull request (No. 45305, per @mylifcc) referenced moonshotai/Kimi-K3.1 and was closed about seven minutes after it opened. The related commit, titled "Kimi-K3.1 support for the K3 chat encoder," is still in the repository's history. It adds K3.1 support to SGLang's experimental router, with 107 test cases. According to @Chinazhidx's readout, K3.1 uses the same tokenizer and chat format as K3: the same tiktoken.model and the same 16 special tokens. The tests add a fourth thinking-effort level, "medium," so the levels become low, medium, high, and max, with max still the default. The tests also cover image input, tool calls, and chat formatting. They check encoding compatibility only, not model quality, so they say nothing about K3.1's performance.

The same test data reportedly shows a live Kimi-K3.1 deployment on DeepInfra on October 18. @NFT_Chen adds that the latest checkpoint is tagged "1019" and that a coding variant is rumored. That chain comes from a single source and has not been confirmed. Separately, a k3d1 identifier, which observers read as K3.1, and a "make3d" mode for Kimi's agent appeared in Moonshot's own web JavaScript. A k3d1-agent string was also spotted on Kimi's website.

Using the same tokenizer as K3 suggests K3.1 is built from K3 and should be cheap for existing K3 integrations to adopt. It does not settle the model's size or architecture. Moonshot has still made no official statement. Community posts expect an anonymous Arena debut under a "kiwi"-style codename, following earlier Kimi models such as K3's "Kivine," before a public launch. That timing is speculation.

Try related models on Kie.ai: Kimi K3, Kimi K3.1 Pro and Claude Sonnet 5.5.