Kimi K3 is the latest large language model from Moonshot AI, a Beijing-based AI lab backed by Alibaba. Released in mid-2026, it is a reasoning-focused model that has drawn attention for benchmark performance and aggressive pricing — sitting alongside DeepSeek as one of the strongest cost-efficient models from China's AI ecosystem.
| Benchmark | Kimi K3 | GPT-4o | Claude Sonnet 4.6 | DeepSeek-V3 |
|---|---|---|---|---|
| MMLU | ~88% | 88% | 90% | 88% |
| HumanEval | ~82% | 90% | 87% | 83% |
| MATH | ~78% | 76% | 78% | 75% |
| Context window | 128K | 128K | 200K | 128K |
| Model | Input per 1M tokens | Output per 1M tokens |
|---|---|---|
| Kimi K3 | ~$0.60 | ~$2.50 |
| DeepSeek-V3 | $0.27 | $1.10 |
| GPT-4o | $5.00 | $15.00 |
| Claude Sonnet 4.6 | $3.00 | $15.00 |
| Gemini 2.5 Flash | $0.10 | $0.40 |
Kimi K3 sits between DeepSeek-V3 and Gemini 2.5 Flash on cost. For teams with data residency concerns about Chinese APIs, consider running open-weight alternatives locally via Ollama or vLLM.
Alongside DeepSeek and Qwen, Kimi K3 is part of a wave of high-capability, low-cost models from Chinese labs reshaping price expectations across the industry.
The real competition in LLMs is now between Chinese open-weight models and Western frontier labs. The cost gap is forcing prices down across the board.
Kimi K3's benchmark profile is strongest on reasoning and math — the areas Moonshot AI explicitly optimized for. Here's how it performs across the standard evaluation suite:
| Benchmark | Kimi K3 | GPT-4o | Claude Sonnet 4.6 | DeepSeek-V3 | Gemini 2.5 Flash |
|---|---|---|---|---|---|
| MMLU | ~88% | 88% | 90% | 88% | ~86% |
| HumanEval (coding) | ~82% | 90% | 87% | 83% | ~80% |
| MATH | ~78% | 76% | 78% | 75% | ~74% |
| GPQA (graduate reasoning) | ~58% | 53% | 65% | 59% | ~55% |
| Context window | 128K | 128K | 200K | 128K | 1M |
The standout result is GPQA — Kimi K3 scores above GPT-4o on graduate-level science and reasoning questions, which reflects Moonshot's investment in chain-of-thought reasoning during training. On pure coding (HumanEval), it trails GPT-4o and Claude Sonnet by a meaningful margin.
Kimi K3 is not fully open source. The model weights are not publicly available for download. Access is through Moonshot AI's hosted API (platform.moonshot.cn) only.
This puts Kimi K3 in the same category as GPT-4o and Claude — a proprietary API model, not a model you can run locally. If you need a comparable open-weight model you can self-host, the closest alternatives are:
Moonshot AI also offers Kimi Code — a coding-specialized variant of the Kimi model family. It's designed for software development tasks: code generation, debugging, code review, and technical documentation.
Kimi Code is available through the same API and targets the same market as GitHub Copilot and Cursor — but as an API-first product rather than an IDE plugin. Early benchmarks show Kimi Code competitive with GPT-4o on HumanEval while being significantly cheaper per token.
For teams building coding assistants or AI-powered development tools, Kimi Code is worth evaluating alongside open-source coding models like Qwen2.5-Coder.
Kimi K3 is available through the Moonshot AI API at platform.moonshot.cn. Pricing as of July 2026:
| Model | Input per 1M | Output per 1M | Context |
|---|---|---|---|
| kimi-k3 (standard) | ~$0.60 | ~$2.50 | 128K |
| kimi-k3 (thinking mode) | ~$0.60 | ~$4.00 | 128K |
| kimi-k3-turbo | ~$0.15 | ~$0.60 | 128K |
Kimi K3 Turbo is the cost-optimized variant — 4x cheaper on input, slightly lower quality. For high-volume production workloads, Turbo is worth benchmarking against the standard model on your specific task. Compared to the field: standard K3 is 4x cheaper than GPT-4o input but 4x more expensive than DeepSeek V4 Flash ($0.14/1M). Turbo is competitive with DeepSeek on input pricing.
Tokoscope works with any OpenAI-compatible endpoint. Free to start.
Get started free →