Kimi K3 with LangChain, LlamaIndex and Claude Code: What Works and What Does Not

Kimi K3 sits at the top of Moonshot's open line: 2,779,931,837,184 total parameters in a Mixture-of-Experts architecture with 93 layers, 896 experts, a 1048576 position limit, and a vision_config plus image_placeholder in the official config.json. The HuggingFace tag is image-text-to-text and the model recorded 751,953 downloads and 11,965 likes in the 30 days sampled in November 2026. K2.5 and K2.6 are earlier 1,026,879,376,368-parameter snapshots with 61 layers and a 262144 context. Only kimi-k3 is on the EXAIV cloud catalog, so K2.5 and K2.6 are local-only and this guide never invents repository names for them.

Specification table

Verified facts versus claims

Every number in this guide comes from the HuggingFace API or the raw config.json of the corresponding repository. That distinction matters because the multimodal story is easy to overstate: the presence of vision_config and image_placeholder confirms an image-capable conditional-generation architecture, but it does not by itself prove video generation, a specific request payload, or a particular benchmark score. Anything beyond the config is marked as needing measurement, and the official model card is the final source.

Memory and deployment reality

Memory planning follows the same rule as any other trillion-parameter MoE: compute scales with active parameters while memory scales with total parameters. A 1T model at 4-bit still needs roughly 500GB for weights alone, which places K2.5 and K2.6 firmly in multi-GPU server territory. Kimi K3 at 2.78T is larger still. For a single consumer card, look at the lightweight families instead. The practical local path is vLLM or SGLang with the official weights, and community conversions should be checked for source and quantization label before use.

Cloud endpoint and cost

On EXAIV, kimi-k3 is callable through the OpenAI-compatible endpoint at https://exaiv.com/api/ai/v1. Integration changes only base_url, api_key and model; streaming and tool calling keep the standard OpenAI format. Prices are never hard-coded here because the live /pricing page is authoritative, and the live /models catalog decides which IDs are actually available. Keep a log of model ID, token counts and latency so that any later model swap can be compared against the same baseline.

Scenario selection

FAQ

Related reading

• Kimi K3 API guide — first API call
• Kimi K3 long context guide — 1M context practice
• Kimi K3 open weight deployment — local weights
• Kimi K3 tool integration — editor and platform setup
• hy4 vs Kimi — lightweight versus flagship

Sources

• moonshotai on HuggingFace
• Kimi-K3 config.json

Parameter and file data sampled in November 2026 and subject to change.