Kimi K3 API 2026: 2.8T Parameters, 1M Context, and CodeFast
Explore Moonshot AI's 2.8T-parameter Kimi K3 with its 1M context, native multimodality, benchmarks, limitations, pricing, and CodeFast API access.
ยท CodeFast Team
What is Kimi K3?
Kimi K3 is the most capable Kimi model Moonshot AI has released to date, announced on July 17, 2026. It combines Kimi Delta Attention (KDA), Attention Residuals (AttnRes), and Stable LatentMoE. Despite its 2.8T total parameter count, it activates only 16 of 896 experts per step. Moonshot says the architecture and training recipe provide roughly 2.5x better overall scaling efficiency than Kimi K2.
Calling K3 merely a larger Kimi misses the point. Its real position is not one-shot code generation but sustained task chains: navigating large repositories, coordinating terminal tools, and incorporating screenshots, images, and video feedback into the workflow. That makes the Kimi K3 API especially relevant to teams building coding agents, research agents, and multi-step knowledge-work systems.
What do the Kimi K3 benchmarks show?
Benchmark Kimi K3 Compared-set position
DeepSWE 67.5 3rd
Terminal-Bench 2.1 88.3 2nd (GPT-5.6 Sol: 88.8)
Program Bench 77.8 1st
SWE Marathon 42.0 1st
Automation Bench 30.8 1st
SpreadsheetBench 2 34.8 1st
BrowseComp 91.2 1st
CharXiv (RQ) w/tool 91.3 2nd
Selected Kimi K3 results from Moonshot AI's official comparison
The table shows K3 pushing the open-model frontier most clearly in terminal work, long software-engineering tasks, and browsing agents. It leads the compared set with 77.8 on Program Bench, 42.0 on SWE Marathon, and 91.2 on BrowseComp. Its 88.3 on Terminal-Bench 2.1 is also very close to GPT-5.6 Sol's 88.8.
Why do 1M context and native multimodality matter?
A 1-million-token context window is not a quality guarantee by itself, but it creates substantial working space for large codebases, long technical documents, multi-turn agent history, and multiple data sources in one task. K3's native visual capabilities also support vision-in-the-loop workflows such as frontend development, game creation, CAD, document analysis, and screenshot-driven debugging.
Known Kimi K3 limitations
- Thinking history must be preserved: quality can become unstable if the agent client does not return the complete prior assistant message and reasoning history.
- Switching from another model to K3 mid-session is not recommended; a fresh session is safer because K3 expects preserved thinking history.
- It can be overly proactive: define boundaries clearly in the system prompt and AGENTS.md so it does not make unexpected decisions on the user's behalf.
- Moonshot itself notes that overall user experience still trails Claude Fable 5 and GPT-5.6 Sol.
Kimi K3 API pricing: official API vs CodeFast
Token type Official Kimi API CodeFast Open Source API
Cache-hit input $0.30 / 1M $0.24 / 1M
Cache-miss input $3.00 / 1M $0.24 / 1M
Output $15.00 / 1M $0.24 / 1M
Cache write $3.00 / 1M $0.24 / 1M
Kimi K3 prices per 1M tokens as of July 18, 2026
CodeFast uses one price for every Kimi K3 token type. That is 12.5x lower than the official cache-miss input price and 62.5x lower than the official output price; the difference for cache-hit input is 1.25x. Actual task cost still depends on reasoning and output token volume. Check the live package page before production because pricing and exchange-rate details may change.
How to use Kimi K3 through the CodeFast API
The CodeFast Open Source API supports the OpenAI-compatible chat/completions format. In an existing OpenAI SDK, Cursor custom provider, or compatible agent client, change the base URL and select kimi-k3 as the model. The shorter codefastk3 alias is also supported for Cursor.
curl https://api.codefast.app/open-source-api/v1/chat/completions \
-H "Authorization: Bearer YOUR_CODEFAST_API_KEY" \
-H "Content-Type: application/json" \
-d '{"model":"kimi-k3","messages":[{"role":"user","content":"Analyze this repository and propose a safe refactor plan."}],"stream":true}'
Basic streaming Kimi K3 request through CodeFast
Kimi K3 or Kimi K2.7 Code?
Choose Kimi K3 for very long tasks, multimodal feedback, large-repository analysis, research, and the highest open-model capability. Kimi K2.7 Code can be more economical for narrower coding tasks and cost-sensitive high-volume workflows; its CodeFast price is $0.108 / 1M tokens. A practical routing strategy sends hard agent jobs to K3 and routine code transformations to a cheaper model such as K2.7 Code.
Is Kimi K3 really open source?
Moonshot describes K3 as the world's first open 3T-class model. However, on this article's publication date, July 18, 2026, the full weights have not yet been released; the announced date is July 27, 2026. Access is available today through Kimi.com, Kimi Code, and the API, but self-hosting and commercial redistribution should not be assumed until the weights and license terms are published.
Verdict: who should use Kimi K3?
- Teams running hours-long coding-agent jobs across large repositories.
- Developers using screenshots, video, or visual feedback for frontend, games, CAD, and document workflows.
- Products that need to combine long documents, research, and multiple data sources inside a 1M context window.
- CodeFast users who want a strong open-model API at a low unit cost without replacing their OpenAI-compatible client.
Kimi K3's main strength is not just its 2.8T parameter count; it is the ability to combine code, terminal work, research, and visual feedback inside one long-running agent loop. Preserved thinking history and boundaries around proactive behavior must be taken seriously in integration. CodeFast offers a practical access layer for teams that want a single predictable token price while retaining an OpenAI-compatible workflow.
Kimi K3 official technical blog
https://www.kimi.com/blog/kimi-k3
Kimi K3 official API quickstart
https://platform.kimi.ai/docs/guide/kimi-k3-quickstart
CodeFast Open Source API package and live pricing
https://codefast.app/packages/open-source-api
Primary sources used in this guide