Kimi-K2.5

Checkpointmoonshotai/Kimi-K2.5Model card indexedModel card

Support summary

Kimi-K2.5 is a checkpoint published by Moonshot AI, built on the KimiK25ForConditionalGeneration architecture. vLLM 0.28.0, the newest indexed vLLM release, documents support. Official recipes or cookbooks cover SGLang, without pinning a stable release.

vLLM documents support in 20 of 20 indexed releases, from 0.17.0 to 0.28.0.

Documented hardware
NVIDIA / CUDA and AMD / ROCm
Documented quantization
int4
Documented features
Reasoning parsers, Speculative decoding, Tensor parallelism, and Tool calling
Configured context
262,144 tokens
License
modified-mit

Hardware, quantization and features count only statements about this checkpoint; engine-wide capabilities are listed with each release.

Hugging Face model card
Publisher
Moonshot AI
Task
image text to text
License
modified-mit
Library
transformers
Architecture
KimiK25ForConditionalGeneration
Weight tensor elements
1T
Tensor types
F32, I32, BF16
Quantization
compressed-tensors · pack-quantized
Configured context
262,144 tokens
Weight access
Public
Repository created
2026-01-01
Source 1: moonshotai/Kimi-K2.5 · publisher model cardSource 2: moonshotai/Kimi-K2.5 · Hugging Face metadataSource 3: moonshotai/Kimi-K2.5 · publisher configurationReviewed
Model notes and sources

Engine support

3 engines

Engine support for Kimi-K2.5
EngineReleaseCheckpoint supportEvidence scopeLast known supported
vLLM0.28.0Newest indexedCheckpoint evidenceSource: vLLM v0.28.0 · docs/models/supported_models.mdReviewed 0.28.0This release
Features, dependencies and evidence for vLLM 0.28.0
SGLangVersion not pinnedRecipe / cookbookDocumented recipe

Official recipe

Source 1: sglang · docs/cookbook/autoregressive/Moonshotai/Kimi-K2.5.mdxSource 2: sglang · docs/src/snippets/autoregressive/kimi-k25-deployment.jsxReviewed
No supported release indexed
Features, dependencies and evidence for SGLang 0.5.19
TensorRT-LLM1.2.1Newest indexedNo matching evidence curatedNo supported release indexed
Features, dependencies and evidence for TensorRT-LLM 1.2.1

Recipes and cookbooks