SearchBook a call

BULLETIN / Moonshot AI

Kimi K3: review

Moonshot's Kimi K3 is among the strongest open-weight models for coding and research, but it is huge, slow and pricey on output.

A top open-weight model for coding and research, but too big to host cheaply and costly through the API.

Maker
Moonshot AI
Released
16 Jul 2026
Licence
Kimi K3 License
Access
Kimi API, Kimi app and web, Kimi Code, open weights on Hugging Face
Context
1,048,576 tokens
Price
$3.00 input (cache miss) / $15.00 output per million tokens (USD)

Good at

  • Web development and coding
  • Text, image and video input
  • Long context of 1M tokens

Bad at

  • Output price close to closed frontier models
  • Slow output and long run times
  • Very large to self-host
  • Custom licence with revenue conditions

Use it for

  • Organisations that need frontier-class open weights
  • Research and coding agents run on own hardware
  • Users of the Kimi Code tool

Avoid it for

  • Low-latency products
  • Budget API use
  • Small teams without multi-GPU servers

Yes, if you need a frontier-class model whose weights you control. Kimi K3 is one of the three highest-scoring open-weight models on the Artificial Analysis index and does well at coding. Through Moonshot’s application programming interface (API), though, its output price is close to closed models, it is slow, and at 2.8T parameters it is out of reach for most self-hosters.

First principles

Open weights help only if you can run them. Downloading the weights gives you control of the model, but only once you have the hardware to serve it. For Kimi K3 that hardware is the main cost of control.

Memory follows total parameters. Each token uses 104B parameters, but all 2.8T must be loaded. Even at 4 bits per parameter (the model card ships a 4-bit format), that is about 1.4 terabytes of weights. One server with eight 80GB graphics processing units (GPUs) holds 640GB, so the weights alone need more than two such servers.

A bill has two sides. Cost is input tokens times the input price plus output tokens times the output price. Caching cuts repeated input to $0.30 per million tokens, so on long agent runs that resend the same context, the $15 output price is the line to watch.

Slow output adds up. Run time is roughly the tokens written divided by output speed. At about 39 tokens per second, long coding and research tasks take a long time, which fits the long Vals run times reported below.

What it is

Kimi K3 is Moonshot AI’s flagship. Fortune reported its launch on 16 July 2026, and Moonshot’s changelog lists it in July 2026. It is a mixture-of-experts model with 2.8T total parameters, of which 104B are active for each token. It reads text, images and video and returns text, with a context window of 1,048,576 tokens.

Moonshot describes it as built for long-horizon coding and knowledge work. Its API offers Chat Completions, Responses and Anthropic-compatible Messages interfaces. The model card recommends vLLM, SGLang or TokenSpeed for serving and ships weights in a 4-bit floating-point format. Moonshot also points users to its Kimi Code command-line agent.

Cheaper siblings remain on the API. kimi-k2.7-code costs $0.95 input and $4.00 output per million tokens, with a 262,144 token window.

Where it is good

Among open models it is near the top of the Artificial Analysis index. It ranks lower on the Arena boards, which mix open and closed models, and two open models score above it on Vals.

  • Artificial Analysis gave it 44 at max effort. On its open-weights page, only MiMo-V2.6-Pro (46) and GLM-5.3 (45) scored higher.
  • On the Arena web development board (updated 29 September 2026) it ranks tenth at 1659, and on the Arena text board (updated 25 September 2026) sixteenth at 1488.
  • ARC Prize lists 94.5% and 60.4% on versions 1 and 2 of the Abstraction and Reasoning Corpus for Artificial General Intelligence (ARC-AGI-1 and ARC-AGI-2).

Moonshot reports 91.2 on BrowseComp and 67.5 on DeepSWE. These are vendor-reported.

Prompt caching makes repeated context cheap: cache-hit input costs $0.30 per million tokens.

Where it is bad

Output price. At $15 per million output tokens, Kimi K3 costs more on output than GPT-6.1 Sol ($10) and Grok 4.7 ($6), and three quarters of what Claude Opus 5.5 charges ($20).

Speed. Artificial Analysis measured about 39 output tokens per second. On the Vals Index (updated 29 September 2026) its average run took 1 hour 32 minutes, one of the longest on the board.

Agentic results. The same Vals Index gives it 50.70% at $6.97 per test. DeepSeek-V4.1-Flash scored slightly higher at $0.33, and MiMo-V2.6-Pro scored 55.20% at $0.41.

Size. With 2.8T parameters, self-hosting needs a cluster of servers with many GPUs. The weights are open, but not practical for a laptop or single-GPU box.

Licence. The Kimi K3 License is custom. Products above 100M monthly active users or $20M monthly revenue must display the model name, and model-as-a-service businesses above $20M revenue over 12 months need a separate agreement.

Billing. Since September 2026, Moonshot’s changelog says account usage is charged half in cash and half in vouchers. Check how that affects your costs.

Price and access

Model Cache-hit input Cache-miss input Output
kimi-k3 $0.30 $3.00 $15.00
kimi-k2.7-code $0.19 $0.95 $4.00
kimi-k2.6 $0.16 $0.95 $4.00

Prices are per million tokens in US dollars, before tax. Moonshot retired kimi-k2.5 and all moonshot-v1 models in August 2026.

Who should use it

On this evidence Kimi K3 is for organisations that want near-frontier coding and research capability on hardware they control, and have the servers to run it. API users chasing value will get more per dollar from DeepSeek-V4.1-Flash, MiMo-V2.6 or GPT-6.1 Sol. It is a poor fit for anything interactive where speed matters.

Sources