Moonshot AI

Kimi K3.1 Pro

Released Oct 23, 2026 · knowledge cutoff unpublished

Status: Unknown
Location: China
Type: LLM
Modality: Multimodal
Context window: 1M
Max output: —
Speed: 52 tok/s
Price ($/MTok in / out): $2 / $10
Cost per task: $2.11
Open weights: Yes

News2 stories

Nov 4, 2026
Moonshot AI publishes the Kimi K3.1 Pro weights on Hugging Face
Oct 23, 2026
Moonshot AI releases Kimi K3.1 Pro, a post-trained Kimi K3 with weights promised by November 4

Strengths & weaknesses

Strengths

Noticeably quicker to answer than K3, whose slowness was its main complaint — early users report close to twice the speed on their own workloads
Built for long agentic runs — Moonshot's demos have it iterating on game builds, CAD scripts and a physics simulator over many steps rather than stopping after a first pass
Fewer invented facts than K3 on Moonshot's own hallucination stress test, a gain Artificial Analysis' independent hallucination eval also shows
Weights are downloadable under the same MIT-derived terms as K3 — ordinary commercial use needs no separate deal, and only the largest hosted services and consumer products pick up conditions

Weaknesses

Thinking is always on and the API rejects requests that disable it; effort defaults to max and changing it mid-conversation throws away the prompt cache
Very verbose — it burns through output tokens, and one subscriber on the $99/month plan hit their five-hour quota in 47 minutes
Launch claims need independent checking — most coding scores were run inside Claude Code, the API docs shipped with placeholder numbers, and allegations of undisclosed paid influencer promotion surfaced in launch week

A post-trained version of Kimi K3 and the first model in Moonshot’s K3.1 series, released 2026-10-23 (2026-10-24 China time) on kimi.ai, the Kimi apps, Kimi Code and the Kimi API; the Hugging Face card gives 2.8T total parameters and 104B active, with MXFP4 weights and MXFP8 activations from quantization-aware training. Weights went public on 2026-11-04 at huggingface.co/moonshotai/Kimi-K3.1-Pro, ungated, under the bespoke Kimi K3.1-Pro License: it starts from MIT wording, then requires a separate agreement for Model-as-a-Service operators above $20M in revenue over 12 months, and a visible "Kimi K3.1-Pro" credit in products above 100M MAU or $20M monthly revenue. Internal use and Moonshot’s certified partners are exempt, so the licence is classed restricted with commercial use permitted. The $2/$10 base price ($0.20 cached input) is relayed by VentureBeat and Handy AI because the official pricing table didn’t render; a kimi-k3.1-pro-highspeed variant costs $4/$20 and Moonshot says it runs 2–3× faster. GPQA Diamond 94.2 and HLE 51.1% (no tools; 70.0% with tools) are Moonshot’s own model-card figures at max effort, and Artificial Analysis measures HLE at 52% no tools. Terminal-Bench 4.0 is 65.7% per Moonshot and 67% per AA, so it can’t be compared with the Terminal-Bench 2.1 field, and BrowseComp (92.1) isn’t tracked. GDPval-AA is 1846 Elo on v2.1, a different scale from the tracked v2, so it’s left null. costPerTask ($2.11, AA’s live page; AA’s launch report said $1.98) and speed are AA’s reasoning variant, which AA doesn’t label with an effort level.

Benchmarks2 of 10 reported

GPQA Diamond

94.2%

Humanity's Last Exam

51.1%

MMLU-Proretired

—

SWE-bench Verified

—

SWE-bench Pro

—

Terminal-Bench 2.1

—

AIME (latest)retired

—

LMArena Elo

—

GDPval-AA v2

—

ARC-AGI-2

—

For developers

API model strings

moonshotkimi-k3.1-pro

Licence

Kimi K3.1-Pro License · Restricted · commercial use permitted

Retirement

No retirement announced

Lineage

No recorded predecessor