Published:
Kimi K3 from Moonshot AI landed on Chat-O today, and it is the largest open-weight reasoning model we have ever offered. At 2.8 trillion parameters with a 1 million token context window, K3 redefines what is possible in a single AI conversation.
Parameter count is not everything, but at this scale it enables emergent behaviors that smaller models cannot replicate. Kimi K3 exhibits a form of deep contextual persistence that allows it to hold complex multi-step reasoning across extremely long conversations without degrading.
| Specification | Kimi K3 | Kimi K2.5 | Industry Average (Power Tier) |
|---|---|---|---|
| Parameters | 2.8T | ~100B | ~200B - 1T |
| Context Window | 1,048,576 tokens | 262,144 tokens | 128K - 200K |
| Modality | Text + Image | Text + Image | Varies |
| Architecture | Open-Weight MoE | MoE | Mostly Closed |
| Release Date | July 16, 2026 | Early 2026 | N/A |
| Tier on Chat-O | Power | Balanced / Power | N/A |
K3 can ingest an entire codebase, understand its architectural patterns, and produce multi-file changes that respect existing conventions. We tested it on a 50,000 line Rust project where it identified a subtle deadlock pattern spanning 12 files and produced a fix across 8 files in a single response.
When operating as the reasoning engine behind an AI agent, K3 maintains coherence across 100+ tool call sequences. Most models degrade after 20-30 steps. K3 stays sharp past 100. This makes it ideal for automated bug hunting, multi-step data pipelines, and complex build systems.
Load a 10,000 page compliance document, ask nuanced questions about cross-referenced clauses, and K3 will answer with citations to specific sections. The 1M token window means you never have to split your source material across multiple chats.
| Model | Parameters | Context | Strengths |
|---|---|---|---|
| Kimi K3 | 2.8T | 1M | Long-horizon reasoning, agentic workflows, multimodal |
| GPT-5 | Undisclosed | 256K | General purpose, broad knowledge |
| Claude Opus 4.6 | Undisclosed | 200K | Safety, nuanced对话 |
| DeepSeek R1 | ~700B | 128K | Math, code reasoning |
| GLM-5.2 | Undisclosed | 1M | Agent workflows, text reasoning |
Kimi K3 separates itself through the combination of open-weight availability, massive parameter count, and the largest context window in its class.
We ran Kimi K3 against a real-world scenario: migrating a Django monolith to a microservices architecture. The model received the full models.py, views.py, urls.py, and settings.py files (roughly 180,000 tokens). It produced:
The entire output was coherent, internally consistent, and accounted for edge cases in the existing authentication flow. No other model we tested could maintain this level of consistency across the full output.
Kimi K3 is available now in the Power tier on Chat-O. Power tier credits apply, and you can switch to K3 from the model dropdown in any chat. There is no waiting list, no API key configuration, and no additional subscription.
The model routes through OpenRouter with provider sort set to price, ensuring you always get the most cost-effective inference backend.
Use Kimi K3 when:
Use something else when:
Kimi K3 is live in the model dropdown. New users get 1,000 free credits to try it. Existing users can select it from the model picker in any chat session. No configuration, no API keys, no waiting.
This is the largest open-weight reasoning model we have ever put behind a chat interface. It changes what you can accomplish in a single session. Sources: