Kimi K2 is an open-weight large language model developed by Moonshot AI, a Beijing-based AI lab. It launched in July 2025 and quickly ranked first among open-source models on the LMSYS Arena leaderboard. The model is optimized for agentic tasks: tool use, multi-step planning, and autonomous code execution.
The easiest way to run Kimi K2 with Ollama is the :cloud tag, a single command that works on any machine with no GPU or download required. For private self-hosted deployments, the second half of this guide covers that path on a Thunder Compute GPU cluster.
Kimi K2 Model Versions
Since the original July 2025 release, Moonshot AI has shipped multiple iterations of the K2 series:
| Version | Release Date | Ollama Tag | Key Addition |
|---|---|---|---|
| Kimi K2 Base | July 10, 2025 | — | Original open-weight release, 128K context |
| Kimi K2 Instruct | September 5, 2025 | kimi-k2 |
Improved agentic coding, 256K context |
| Kimi K2 Thinking | November 6, 2025 | Retired June 16, 2026 | Extended chain-of-thought reasoning mode |
| Kimi K2.5 | January 27, 2026 | kimi-k2.5:cloud |
Native multimodal (vision + text), Agent Swarm |
| Kimi K2.6 | April 20, 2026 | kimi-k2.6:cloud |
300-agent swarm, production-grade coding, multimodal |
| Kimi K2.7 Code | June 12, 2026 | kimi-k2.7-code:cloud |
Coding-focused, ~30% fewer thinking tokens vs K2.6 |
For most workloads, kimi-k2.6:cloud is the right tag for general chat, multimodal input, and agent orchestration. For agentic coding specifically, kimi-k2.7-code:cloud cuts thinking-token usage by roughly 30% with the same 1T-parameter architecture.
Understanding the Kimi K2 Model
Kimi K2 uses a Mixture-of-Experts (MoE) architecture, which lets it scale to massive parameter counts without proportionally increasing inference cost. A routing mechanism selects a small subset of specialized "expert" networks for each input rather than activating every parameter. This is how K2 carries 1T total parameters while computing through only 32B per forward pass.
The model was trained on 15.5T tokens using the MuonClip optimizer, a method Moonshot AI developed to stabilize large-scale MoE training. The instruct variant is the right choice for chat, agentic use, and tool-calling workflows.
Kimi K2 Context Window Size
In July 2025, the original K2 Base shipped with a 128K context window. The Instruct version extended this to 256K tokens, and K2.5, K2.6, and K2.7 Code all maintain the 256K window. For self-hosted inference with llama.cpp or Ollama, context window size directly affects VRAM requirements, meaning it's best to configure it conservatively unless you have ample memory headroom.
Kimi K2 Benchmarks: How It Compares to GPT and Claude
The numbers below are from the official technical report and third-party evaluations, all under non-thinking (standard inference) settings.
| Benchmark | Kimi K2.6 | GPT-5.5 | Claude Opus 4.7 |
|---|---|---|---|
| SWE-Bench Verified | 80.2% | 88.7%1 | 87.6% |
| SWE-Bench Pro | 58.6% | 58.6%2 | 64.3% |
| LiveCodeBench v6 | 89.6% | — | — |
| BrowseComp | 83.2% / 86.3%3 | 84.4% | 79.3% |
| HLE with tools | 54.0% | 41.4% | 54.7% |
For a broader comparison of leading open-source models, see the Thunder Compute guide to the best open-source LLMs.
How Good Is Kimi K2?
Kimi K2 Strengths
Kimi K2 is genuinely strong in the areas it was designed for. SWE-Bench results show it consistently outperforms GPT-4.1 and DeepSeek-V3 on software engineering tasks. On LiveCodeBench, which tests real coding scenarios that can't be memorized from training data, K2 leads the open-source field by a meaningful margin.
Kimi K2 Weaknesses
The base K2 model is text-only; vision capabilities were added in K2.5 (January 2026). Inference speed on quantized local builds is a real constraint: the 1.8-bit quant maxes out at a few tokens per second without enterprise multi-GPU hardware.
Kimi K2 Use Cases
K2 is best suited for:
- Developers who need a high-quality open-weight coding and agentic model
- Teams that want to self-host for data privacy reasons
- Workflows that can route inference through cloud GPUs like Thunder Compute when local hardware falls short
How to Run Kimi K2 with Ollama in Cloud Mode
Ollama's :cloud tags send prompts to Moonshot's infrastructure through Ollama's servers, streaming responses back to your terminal exactly like a local model. Only a small manifest file (a few KB) is stored on your machine.
Step 1: Install Ollama
Download and install Ollama from ollama.com for macOS, Windows, or Linux. On macOS, brew install ollama also works.
Step 2: Sign In
Cloud models require an Ollama account. Sign in from the terminal:
ollama signin
This opens a browser prompt. Complete the approval and return to the terminal. As of July 2026, signing in does not require payment information. Basic usage of cloud models is free within Ollama's usage limits.
Step 3: Run Kimi K2.6
ollama run kimi-k2.6:cloud
The first response can take 10–30 seconds while Ollama establishes the cloud session. After that, responses stream at normal speed. ollama list will show kimi-k2.6:cloud at only a few KB — expected, since inference runs on Moonshot's servers.
For agentic coding specifically, kimi-k2.7-code:cloud offers roughly 30% fewer thinking tokens per task:
ollama run kimi-k2.7-code:cloud
Step 4: Use K2.6 with Coding Agents
Ollama supports launching popular coding agents with Kimi K2.6 as the backend:
# Claude Code
ollama launch claude --model kimi-k2.6:cloud
# OpenCode
ollama launch opencode --model kimi-k2.6:cloud
# Codex App
ollama launch codex-app --model kimi-k2.6:cloud
Important: The original cloud tags
kimi-k2:1t-cloudandkimi-k2-thinkingretired on June 16, 2026. Replace them withkimi-k2.6:cloudin any scripts, Modelfiles, or agent configs, or requests will fail.
Why Self-Hosting Kimi K2 Requires a Cloud GPU
For private inference with regulated data, very high sustained throughput, or custom fine-tuned weights, the cloud tag won't work and you'll need to self-host the model weights. The hardware requirements are substantial.
The Hardware Requirements at 1T Parameters
K2's MoE architecture keeps inference memory requirements lower than a dense 1T model, but they're still significant. Model weights in FP8 format take up about 1TB on disk. Even the 1.8-bit quantized GGUF comes in around 250GB, requiring at least 247GB of combined RAM and VRAM for usable throughput.
A Q4 quantization runs to approximately 584GB, requiring at least 600GB of combined RAM and VRAM, meaning multi-GPU server configurations like 8x H100 or H200 nodes. A 24GB consumer GPU like an RTX 4090 can handle the 1.8-bit variant by offloading MoE layers to system RAM, but inference speed drops to roughly 1–2 tokens per second, too slow for practical use.
For a full breakdown of which cloud GPU fits your workload and budget, see the Thunder Compute guide to the best GPU for LLM work.
Why Thunder Compute Is a Viable Option
Thunder Compute is a Y Combinator-backed cloud GPU provider offering on-demand access to A100s and H100s at a fraction of the cost of owning hardware. It ships a pre-configured Ollama template, removing the biggest friction point when working with large models in the cloud.
| Configuration | VRAM | Quantization | Expected Speed | Storage | Hourly Price |
|---|---|---|---|---|---|
| 4× A100 80GB1 | 320 GB | Q2 GGUF (~340 GB, CPU offload required) | 1–5 tok/s | 400GB | $5.96 |
| 4× H100 PCIe 80GB1 | 320 GB | Q2 GGUF (~340 GB, CPU offload required) | 3–8 tok/s3 | 400GB | $11.56 |
| 8× A100 80GB | 640 GB | Native INT4 (~594 GB) | 10–20 tok/s | 700GB | $11.92 |
| 8× H100 PCIe 80GB | 640 GB | Native INT4 (~594 GB) | 25–40 tok/s3 | 700GB | $23.12 |