What is Kimi K2 Thinking good for?

Multi-step reasoning and long tool-calling chains: Moonshot built it as a thinking agent that mixes chain-of-thought with tool calls. Modified MIT weights allow self-hosting, on a multi-GPU node.

When is Kimi K2 Thinking not a good fit?

New deployments: Moonshot retired it in May 2026, and the later Kimi models carry the line on. Text only, and running the weights yourself takes a multi-GPU node.

What is the cheapest way to run Kimi K2 Thinking?

Hosted, unless you push serious volume. Novita charges $0.60 in / $2.50 out per 1M tokens. The cheapest rental that fits is 8x A100 at $10,138 a month, which costs the same as about 11B tokens a month on that API.

Can I self-host Kimi K2 Thinking?

Yes. The weights are Modified MIT licensed. At 4-bit it needs about 554 GB of GPU memory, which starts at roughly $10,138 a month on the cheapest rental that fits.

Retired May 2026 Open weights · Modified MIT

Kimi K2 Thinking

Reasoning variant of Kimi K2, Moonshot AI's 1T-parameter MoE, designed for deep multi-step reasoning and complex analysis.

Jump to: Specs Pricing Self-hosting Similar models FAQ More from Moonshot AI

Visit website Compare vs other LLMs →

Cheapest via Novita

Input: $0.60 /1M tokens
Output: $2.50 /1M tokens

2 providers

Key Specifications

Context window: 262K tokens
Parameters: 1T, 32B active
Inputs: Text

Capabilities Show details: Reasoning Thinks before it answers, always on or as a switchable mode. Function calling Connect to external tools, APIs, and systems. Structured output Return responses in structured formats like JSON. Web search Search the web for up-to-date information.

Hosted API pricing

Input tokens M Output tokens M

Provider	Input /1M tokens	Output /1M tokens	Cached input /1M	Cost at 10M in + 2M out
Novita	$0.60	$2.50	$0.15	$11.00	View
Moonshot AI Creator	Open weights, no hosted price				View

Heads up: Base-tier, on-demand rates per 1M tokens; cached, batch and long-context tiers excluded. A provider may serve a shorter context or a quantized build than the creator's release. Verify before provisioning. More on how we price.

Estimated cost to self-host

Kimi K2 Thinking needs about 554 GB of GPU memory at 4-bit with 32K context. The cheapest rental that fits is 8× A100 at about $10,138 a month.

That costs the same as roughly 11B tokens a month on Novita's API. Self-hosting is more expensive below that volume.

Single node ≤ 8

Precision	Memory	Cheapest fit (32K context)(Full 262K context)	Cost /mo	Break-even vs API
4-bitINT4 / FP4	554 GB570 GB	8× A100	$10,138$10,138	11B tokens /mo11B tokens /mo
8-bitFP8 / INT8	1,003 GB1,011 GB	4× B300	$22,723$22,723	25B tokens /mo25B tokens /mo
16-bitFP16 / BF16	2,004 GB2,020 GB	8× B300	$45,446$45,446	50B tokens /mo50B tokens /mo

Estimates based on median on-demand rates for Nvidia GPUs. Memory is weights plus KV cache for one request, using FP8 KV cache where supported and FP16/BF16 otherwise. Break-even assumes a 5:1 input-to-output ratio. No guarantee of runtime support, usable performance, or that a matching quantized build exists. How we estimate costs.

Similarly priced models

The models nearest Kimi K2 Thinking by blended rate, each at its own cheapest provider.

Model	Blended /1M	Input /1M	Output /1M	Context	Cutoff	vs Kimi K2 Thinking
Qwen3.5-Plus Alibaba Cloud	$0.7333	$0.40	$2.40	1M		−20%
GLM-5 Z.AI	$0.7667	$0.60	$1.60	200K		−16%
DeepSeek R1 Distill Llama 70B DeepSeek	$0.80	$0.80	$0.80	131K		−13%
GLM-4.7 Z.AI	$0.8667	$0.60	$2.20	205K		−5%
Gemini 3 Flash Preview Google Cloud	$0.9167	$0.50	$3.00	1M	Jan 2025	0%
Kimi K2 Thinking This model Moonshot AI	$0.9167	$0.60	$2.50	262K
Qwen3.6-Plus Alibaba Cloud	$0.9167	$0.50	$3.00	1M		0%
GLM-5.2 Z.AI	$0.95	$0.70	$2.20	1M		+4%
Nemotron 3 Ultra 550B A55B Nvidia	$0.99	$0.66	$2.64	1M	Sep 2025	+8%
DeepSeek R1 DeepSeek	$1.00	$0.70	$2.50	128K		+9%
DeepSeek R1 Turbo DeepSeek	$1.00	$0.70	$2.50	64K		+9%

Prices are USD per 1M tokens at each model's cheapest listed provider. Blended is the cost of 10M input plus 2M output tokens, spread over the 12M.

Lium

Built for agents: CLI, SDK and MCP server. Your agent rents and runs GPUs itself.

A permissionless marketplace. 1000+ nodes from independent providers worldwide.

Hard to find nodes available; B200 and B300 Blackwell nodes listed now.

Consumer + DC grade cards for cheap. H100s from $1/hr, RTX 5090s from $0.35/hr.

View Pricing

Frequently Asked Questions

What is Kimi K2 Thinking good for?: Multi-step reasoning and long tool-calling chains: Moonshot built it as a thinking agent that mixes chain-of-thought with tool calls. Modified MIT weights allow self-hosting, on a multi-GPU node.
When is Kimi K2 Thinking not a good fit?: New deployments: Moonshot retired it in May 2026, and the later Kimi models carry the line on. Text only, and running the weights yourself takes a multi-GPU node.
What is the cheapest way to run Kimi K2 Thinking?: Hosted, unless you push serious volume. Novita charges $0.60 in / $2.50 out per 1M tokens. The cheapest rental that fits is 8x A100 at $10,138 a month, which costs the same as about 11B tokens a month on that API.
Can I self-host Kimi K2 Thinking?: Yes. The weights are Modified MIT licensed. At 4-bit it needs about 554 GB of GPU memory, which starts at roughly $10,138 a month on the cheapest rental that fits.

More from Moonshot AI

Model	Context	Input /1M	Output /1M
Kimi K2 Open weights	262K	$0.57	$2.30
Kimi K2.6 Open weights	262K	$0.60	$3.20
Kimi K2.5 Open weights	262K	$0.40	$2.60
Kimi K2.7 Code Open weights	262K	$0.90	$3.75
Kimi K3 Open weights	1M	$2.00	$12.00