📊 See Kimi K3 (Max) on Arena

📱 Get the AI Rank app

Kimi K3 is Moonshot's open-weight 2.8-trillion-parameter flagship (first open model in the 3T class per Moonshot). Searches for kimi 3.1 or kimi k3.1 map to the K3.1 Pro post-training line on top of K3, not a separate base-weight family. API ids on the platform today: kimi-k3, kimi-k3.1-pro, and kimi-k3.1-pro-highspeed.

Kimi K3 open weights and architecture

The MoonshotAI/Kimi-K3 README and the Hugging Face card agree on the headline specs:

  • 2.8T total parameters; Moonshot calls it the first open 3T-class model
  • 1,000,000-token context; native text / image / video understanding
  • Architecture notes: Kimi Delta Attention (KDA), Attention Residuals (AttnRes), MoE activating 16 of 896 experts under Stable LatentMoE
  • Full weights released under the Kimi K3 License (not a generic Apache/MIT badge)
  • Platform access: select model id kimi-k3 on platform.kimi.ai (OpenAI/Anthropic-compatible APIs)

Moonshot’s models list still describes K3 as the previous-generation flagship beside the newer K3.1 Pro SKUs.

K3.1 Pro as post-trained flagship

The Kimi K3.1 Pro blog states that Kimi K3.1 Pro is a post-trained version of Kimi K3 with substantially larger post-training compute, still in the 3T-class open-weight line, with a 1M-token context. Kimi K3.1 Pro Highspeed is the same intelligence at about 2-3x speed. Availability named in the post: Kimi.ai, Kimi Work, Kimi Code, and the Kimi API. Full model weights will be released by Nov 4, 2026 (blog wording).

Vendor-reported scores on that post (not re-run here): Terminal-Bench 4.0 65.7% (blog compares to GPT-6 Astra 58.2% and Claude Opus 5.5 66.4%); Kimi Code Bench 3.0 Long-horizon 61.2% vs K3 55.95%. Treat these as Moonshot’s published figures.

API model ids

From the models catalog and quickstarts:

API id Role
kimi-k3 Open-weight 2.8T K3 base flagship
kimi-k3.1-pro K3.1 series flagship (post-trained K3)
kimi-k3.1-pro-highspeed Same as Pro, ~2-3x faster

Quickstarts use base_url https://api.moonshot.ai/v1 with those model strings. Both K3 and K3.1 Pro always think; reasoning_effort supports low / high / max (default max). Context for the K3 series on the pricing page is 1,048,576 tokens with flat (non-length-tiered) pay-as-you-go rates.

Official platform pricing

Checked on Model Inference Pricing (USD per MTok = 1,000,000 tokens). Columns: Input, Output, Cached Input, 5m cache writes, 1h cache writes.

Model Input Output Cached Input 5m writes 1h writes
kimi-k3.1-pro $2.00 $10.00 $0.20 $2.00 $4.00
kimi-k3.1-pro-highspeed $4.00 $20.00 $0.20 $4.00 $8.00
kimi-k3 $3.00 $15.00 $0.30 $3.00 $6.00

Taxes may apply at checkout. Context caching is automatic for eligible prefixes (default TTL 5m). Flagship models unlock after a successful top-up (minimum $1 per quickstart notes). This is API pay-as-you-go, separate from Kimi membership / Kimi Code plans.

→ Live usage: See OpenRouter token share on the AI Rank models board.

📱 Follow rankings in the AI Rank app

AI Rank snapshot

  • Arena overall (source publish date 2026-10-15): Kimi K3 (Max) rank 18, rating about 1487, 32,795 votes — model page. That row is the Arena id, not the API string kimi-k3.1-pro.
  • OpenRouter week text window 2026-10-10 to 2026-10-16: moonshotai/kimi-k3-20260715 at rank 11 with 2.12T tokens in the AI Rank feed (batch-labelled name on the board).

No separate Arena id named “K3.1 Pro” appeared in the same overall list when checked.

Sources

Compare before you choose: check current model rankings on AI Rank before you commit.

Or 📱 get the AI Rank app to follow them on your phone.

Prepared with automated editorial assistance and checked against Moonshot GitHub/HF, Kimi platform docs/pricing, the K3.1 Pro blog, and dated AI Rank Arena/OpenRouter feeds on October 24, 2026. We did not run our own benchmarks for this article.