What is Kimi K3?
Kimi K3 is a mixture-of-experts model from Moonshot AI with 2.8T parameters, 104B of them active for each token. It reads text, images and video, writes text, and always reasons before it answers; you pick low, high or max effort, and the API defaults to max. Moonshot says it still trails Claude Fable 5 and GPT-5.6 Sol overall but beat every other model it tested.
It is Moonshot's current flagship, in the Kimi app, Kimi Work, Kimi Code and the Kimi API, and it is where Moonshot sends users of its retired Kimi K2 models. The weights are free on Hugging Face, but they take about 1.56 TB.
What changed from Kimi K2?
Almost everything: Kimi K3 is nearly three times the size, reads images and video, holds eight times the context and always reasons first.
- Size
- 2.8T parameters with 104B active per token, up from 1T and 32B.
- Context
- 1M tokens, up from 128K on the original Kimi K2.
- Input
- Text, images and video; Kimi K2 read text only.
- Thinking
- Always on, with low, high and max effort; Kimi K2 answered without a long reasoning phase.
- Licence
- Adds a rule that large model-hosting firms need a separate agreement with Moonshot.
- API
- Moonshot has discontinued every Kimi K2 model name and points those users to Kimi K3.
Who is Kimi K3 best for?
Kimi K3 suits long coding sessions, frontend and 3D work, deep research, and teams that want open weights at the largest scale available.
Long coding sessions
Moonshot built it to work through long engineering jobs with little supervision, and it posted the top SWE-Marathon score in Moonshot's own table, 42.0 against 35.0 for Claude Fable 5.
Frontend, 3D and visual work
It checks screenshots of its own output as it codes, and early users single out its interfaces, animations and 3D models.
Research and knowledge work
It scored 91.2 on BrowseComp in Moonshot's tests and builds research reports with charts and dashboards in Kimi Work.
Teams that need open weights at this scale
Companies with a large GPU cluster can host or fine-tune it; the licence allows commercial use, with extra terms only for very large products and model-hosting firms.
What's included with Kimi K3?
Three effort levels, image and video input, automatic caching across a 1M context, flexible tool calling, and open MXFP4 weights.
- Three effort levels
- Low, high and max trade answer depth against speed and tokens; thinking cannot be switched off.
- Image and video input
- Images go in as base64 and videos as uploaded files; public image URLs are not accepted.
- Automatic context caching
- Repeated prefixes are cached for 5 minutes or 1 hour, and cache hits cost $0.30 per million tokens.
- Tools and structured output
- Forced tool calls, tools added mid-conversation, strict JSON schemas and a partial mode that continues from a prefix.
- Works in coding agents
- Moonshot recommends its own Kimi Code CLI and publishes setup guides for Claude Code, Codex, OpenCode and Hermes Agent.
- Open weights in MXFP4
- Trained with 4-bit weights from the fine-tuning stage on, and served by vLLM, SGLang and TokenSpeed.
How much does Kimi K3 cost?
Kimi K3 costs $3 per million input tokens and $15 per million output tokens on Moonshot's API, the weights are free to download, and Kimi plans that include it in Kimi Code start at $19 a month.
- Full weights on Hugging Face, about 1.56 TB
- Kimi K3 License, commercial use allowed
- Runs on vLLM, SGLang and TokenSpeed
- Cache hits $0.30; cache writes $3 (5 minutes) or $6 (1 hour)
- One flat rate across the 1M context
- Opens after a $1 minimum top-up
- K3 in Kimi Code with a 256K context
- Agent credits for the Kimi apps
- Deep Research, documents, sheets and slides
- Twice the agent credits of Plus
- K3 with the full 1M context in Kimi Code
- Goal mode and one-click Kimi Claw
Kimi Max ($99) and Ultra ($199) a month add more credits and are the only plans with 1M-token K3 chats in the Kimi app. Paying yearly brings Plus, Pro, Max and Ultra to $15, $31, $79 and $159 a month. The free Kimi app plan also lists K3 in its model menu. API prices are in US dollars and exclude tax, and K3 is not on Moonshot's Batch API. Full pricing on platform.kimi.ai. We checked these prices on .
How do readers rate Kimi K3?
No one has rated Kimi K3 yet. Be the first.
Sign in to rate. One rating per account.
Kimi K3 pros and cons
Users rate its coding, visual work and writing highly, but it is slow, token-hungry, expensive for an open model and too large for most teams to host.
Pros
- Early users rate its frontend, animation and 3D work highly, some above what Claude Fable and GPT-5.6 Sol produced on the same jobs.
- Users who put it to real work call its results Opus-level, and writing and roleplay users rank its prose among the best they have tried.
- Moonshot's own table puts it level with Claude Fable 5 on Terminal-Bench 2.1 (88.3 against 88.0) and ahead of it on SWE-Marathon and BrowseComp.
- The first open-weight model to reach 2.8T parameters, by Moonshot's account, under a licence that allows commercial use and fine-tuning.
Cons
- Slow and token-hungry: it reasons at max effort by default, and users report half-hour waits, drained API balances and a $19 plan's five-hour limit used up on one small task.
- Expensive for an open model: $3 / $15 per million tokens is well above what Xiaomi, Z.ai and DeepSeek charge for their open models, and many users say so.
- Fussy to integrate: Moonshot warns it turns unstable if a harness drops its reasoning history or switches to it mid-session, and early users hit timeouts, API errors and runs that stopped early.
- Too big for nearly anyone to self-host: the weights take about 1.56 TB, and Moonshot recommends servers with 64 or more accelerators.
What are the best Kimi K3 alternatives?
Three tools on Recaply do Kimi K3's job. MiMo-V2.6-Pro is open source, so you can run it on your own server.
MiMo-V2.6-ProOpen SourceChoose it when price matters most: Xiaomi charges $0.435 / $0.87 per million tokens, a small fraction of K3's rate, and its MIT-licensed weights also read audio.
GLM-5.3FreeChoose it when you only need text and want a cheaper open coding model: it costs $1.40 / $4.40 per million tokens with the same 1M context.
Claude Opus 5.5PaidChoose it when you would rather pay a third more for a closed model from Anthropic: $4 / $20 per million tokens, with the same 1M context at one flat rate.
Questions about Kimi K3
What is Kimi K3?
Kimi K3 is Moonshot AI's 2.8-trillion-parameter open-weight model for long coding jobs, agents and research, with a 1M-token context, image and video input, and API prices of $3 / $15 per million tokens.
Who is Kimi K3 for?
Kimi K3 suits long coding sessions, frontend and 3D work, deep research, and teams that want open weights at the largest scale available.
Is Kimi K3 free?
Partly. The weights are free to download from Hugging Face, and the free Kimi app lists K3 in its model menu. The Kimi API is paid, from $3 per million input tokens after a $1 minimum top-up, and Kimi Code access starts with the $19 Plus plan.
How much does Kimi K3 cost?
Kimi K3's paid plans are Kimi API at $3 / $15 per million input / output tokens, Kimi Plus at $19 per month and Kimi Pro at $39 per month.
Is there an open source alternative to Kimi K3?
Yes. MiMo-V2.6-Pro is an open source tool that does the same job, so you can run your own.
What are the best Kimi K3 alternatives?
The closest alternatives on Recaply are MiMo-V2.6-Pro, GLM-5.3 and Claude Opus 5.5.
Where can I try Kimi K3?
On its official website, huggingface.co.
Can I run Kimi K3 locally?
Not on ordinary hardware. The weights take about 1.56 TB even in 4-bit MXFP4, and Moonshot recommends servers with 64 or more accelerators. vLLM, SGLang and TokenSpeed can serve it for teams with that kind of cluster.
Can I use Kimi K3 in a commercial product?
Yes, under the Kimi K3 License. Two extra terms apply: firms that sell model access through an API and earn over $20 million in any 12 months need a separate agreement with Moonshot, and products with over 100 million monthly users or $20 million monthly revenue must show "Kimi K3".
Can I turn off Kimi K3's thinking?
No. Kimi K3 always reasons before it answers. To save time and tokens, set reasoning_effort to low instead of the default max, and in multi-turn chats send back the whole assistant message, reasoning included, or its answers can turn unstable.
Which Kimi plans include Kimi K3 in Kimi Code?
Plus at $19 a month and above. Plus runs K3 with a 256K context; Pro at $39 and higher plans add the full 1M context, which uses about twice as much quota as the 256K version.
Our verdict
Kimi K3 is worth testing for hard coding and visual work, but it is slow, costly and hard to host for an open model.
Moonshot's own tests put it close to Claude Fable 5, and early users praise its frontend, 3D and writing results, but it reasons at max effort by default and costs $3 / $15 per million tokens.
- Skip it for
- Budget or high-volume agent work, where MiMo-V2.6-Pro or GLM-5.3 cost a fraction as much, and anyone hoping to run it on their own hardware.
- Check first
- Run your own tasks at low and high effort and watch the token count before you commit to a plan.
Kimi K2.5
Kimi K2.6
Kimi K2
Command A+
LFM3