Kimi K3 review, pricing and verdict

Pick Kimi K3 if you want an open-weight model for agentic coding and long-document work and can budget for multi-GPU hosting or the hosted API. Skip it where verified factual answers, audio or video input, or Western compliance certificates are required, because those are its weak spots.

Kimi K3 is Moonshot AI's mixture-of-experts language model, released July 16, 2026, with native image input and a 1M-token context window built on Kimi Delta Attention. Its weights have been public on Hugging Face since July 27, 2026, under the Kimi K3 License, and it scores 93.5% on GPQA Diamond.

Provider: Moonshot AI · Family: Kimi

More about Moonshot AI on HokAI

Context window: 1,048,576 tokens · Max output: 128,000

Input modalities: text, image · Output: text, tool-calls, code

About Kimi K3

Kimi K3 is a 2.8-trillion-parameter mixture-of-experts language model from Moonshot AI, released July 16, 2026. It routes each token through 16 of 896 experts, roughly 280B active parameters, and uses Kimi Delta Attention, a hybrid linear attention design paired with Attention Residuals that is meant to keep recall stable across very long inputs. It follows the Kimi K2.7 Code HighSpeed line and is positioned as Moonshot's frontier release for agentic coding, long-horizon knowledge work and multimodal reasoning, competing with GPT-5.6 Sol and Claude Opus 4.8.

Moonshot's launch post reports 93.5% on GPQA Diamond, the highest published open-weight result and ahead of Opus 4.8 at 91.0%. On agentic coding it lists 67.5% on DeepSWE (67.3% under the mini-SWE-agent scaffold) and 88.3% on Terminal-Bench 2.1, half a point behind GPT-5.6 Sol, plus 77.8% on Program Bench and 42.0% on SWE Marathon. These are vendor-run scores on Moonshot's own scaffolding, so read gaps of a few points as indicative rather than settled.

The context window is 1,048,576 tokens with a 128,000-token output cap, matching GPT-5.5 and exceeding the 200K windows on the GPT-5.6 Sol, Terra and Luna models. K3 takes text and image input and returns text, tool calls and code. Vision scores include 81.6% on MMMU-Pro and 97.8% on MathVision. It supports function calling, structured output and code execution, but has no native audio or video, so voice and video pipelines need a second model. Reasoning modes run from none to max, and Moonshot has not published the token budget for each level.

Access is through kimi.com, the Kimi iOS app, Kimi Code and the API at api.moonshot.ai (model ID kimi-k3). The same model is routed through OpenRouter. The weights were published on Hugging Face as moonshotai/Kimi-K3 on July 27, 2026, a few days after launch, under the bespoke Kimi K3 License rather than a stock OSI license, so redistribution and fine-tuning terms need reading in the repository. Self-hosting is a serious hardware job: full FP16 inference needs roughly 5.6TB of VRAM, and even INT4 quantization needs about 1.4TB, so most teams will use the hosted API or a provider on OpenRouter.

Moonshot uses RLHF alignment with an internal safety team, a content filter and a moderation API. The independently measured HarmBench refusal rate is 88%. The caution is factual reliability: K3 posts a 51% hallucination rate on AA-Omniscience, up from 39% on K2.6, so it answers more confidently while fabricating more often. Ground factual work in retrieval or add a verification step. Moonshot discloses no SOC 2, HIPAA or ISO 27001 certification, and documentation is Chinese-first. Hosted API requests are retained for 30 days for abuse monitoring, and Moonshot states it does not train on API inputs by default. Training data is described as public web text, licensed datasets, synthetic reasoning traces and internal Chinese and English corpora, with no published training cutoff.

K3 suits teams that want an open-weight frontier model for long-context agentic coding, graduate-level reasoning or fine-tuning research and can live with a Chinese-first ecosystem. Teams that need verified factual reliability, native audio or video, or Western enterprise certificates will do better with Opus 4.8 or GPT-5.6 Sol. GLM-5.2 and DeepSeek V4 Pro are the other open-weight models worth comparing in the directory. For the consumer-chat angle, see Kimi versus ChatGPT, or let Smart Match pick a model for your task.

Pricing

Hosted API (api.moonshot.ai) bills per token: $3.00 per 1M input, $0.30 per 1M cached-hit input (a 90% discount) and $15.00 per 1M output, flat across the full 1M window with no long-context surcharge. OpenRouter carries the official rates under moonshotai/kimi-k3, while independent hosts there now undercut them, with input as low as roughly $1.70 per 1M. Local inference from the open weights costs hardware only.

What a real job costs

Budgets: 20-page PDF = 30k in / 1k out · Support reply = 2k in / 300 out · Coding agent run = 200k in / 20k out. Computed from the vendor's per-token prices at render time; cached-input discounts are not applied.

Key Features

• Mixture-of-experts at frontier scale: 896 experts with only a small active subset per token, so inference cost tracks active rather than total parameters.
• Long context with Delta Attention: Kimi Delta Attention is built to keep recall high across the full window instead of degrading past 100K tokens as sliding-window designs do.
• Native vision on an open model: Text and image input with strong vision scores, a rare combination among open-weight models.
• Open weights under a custom license: Weights sit on Hugging Face under the Kimi K3 License, so private hosting and fine-tuning are possible once the terms are read.
• Six reasoning modes: Reasoning effort runs from none to max, letting teams trade latency and cost against depth per request.

Pros

• Leads published open-weight GPQA Diamond results, which makes it the safest open pick for graduate-level science and reasoning work.
• Terminal-Bench 2.1 at 88.3% sits within a point of GPT-5.6 Sol, a credible result for terminal-driven agents on an open model.
• Program Bench at 77.8% and SWE Marathon at 42.0% top the tested field in Moonshot's reported results, a useful signal for long-running coding agents.
• No long-context surcharge on the hosted API means a full-window request costs the same per token as a short one.
• Public weights give teams a route to private hosting and fine-tuning that closed frontier models do not offer.

Cons

• Hallucination rate rose to 51% on AA-Omniscience versus 39% for K2.6, so factual output needs grounding or review.
• Moonshot paused new consumer subscriptions on kimi.com between July 20 and September 19 during a GPU capacity crunch, so access can tighten when demand spikes.
• Self-hosting means a multi-node GPU cluster, with roughly 5.6TB of VRAM for full precision and about 1.4TB even at INT4.
• No native audio or video input or output; K3 handles text and images only.
• No SOC 2, ISO 27001 or HIPAA disclosure, and the community is Chinese-first.

Benchmarks

• MATH: 94% vendor-reported · 16 Jul 2026 — Competition maths problems, % solved.
• MMLU: 95% vendor-reported · 16 Jul 2026 — General-knowledge exam across 57 subjects, % correct.
• MMLU-Pro: 87% vendor-reported · 16 Jul 2026 — A harder version of the 57-subject knowledge exam, % correct.
• AIME 2025: 89% vendor-reported · 16 Jul 2026 — Competition-level maths problems from the 2025 exam, % solved.
• ARC-AGI 2: 32% vendor-reported · 16 Jul 2026 — Abstract visual puzzles built to resist memorisation, % solved.
• HumanEval: 92% vendor-reported · 16 Jul 2026 — Small programs that must pass hidden tests, % passing.
• LiveBench: 68% vendor-reported · 16 Jul 2026 — A rolling set of fresh questions that cannot have been in training data, % correct.
• LMArena Elo: 1380 independent · 16 Jul 2026 — Rating from blind human votes on which answer is better.
• GPQA Diamond: 93.5% vendor-reported · 16 Jul 2026 — PhD-level science questions that are hard to search for, % correct.
• LMArena rank: #8 independent · 16 Jul 2026 — Position on the blind human-preference leaderboard; #1 is best.
• Aider Polyglot: 77.8% vendor-reported · 16 Jul 2026 — Code edits across several programming languages, % correct.
• SWE-bench Verified: 67.5% independent · 16 Jul 2026 — Real GitHub issues fixed end to end, % solved.
• Humanity's Last Exam: 43.5% vendor-reported · 16 Jul 2026 — Expert-written questions across many fields, % correct.
• AA Intelligence Index: 44 cited: Artificial Analysis · 23 Sep 2026 — Composite of 10 evaluations run by Artificial Analysis, 0 to 100.
• AA blended price: $2.31/M cited: Artificial Analysis · 23 Sep 2026 — Price per 1M tokens at a 3:1 input to output blend, as listed by Artificial Analysis.
• Output speed: 37 tok/s cited: Artificial Analysis · 23 Sep 2026 — Median tokens written per second as measured by Artificial Analysis.

A benchmark is an exam, not the job. Scores transfer unevenly between tasks, so weigh the one closest to your workload and read every figure with its source.

Visit Kimi K3 Official Page