GLM-5.2 is live. #1 throughput on OpenRouter, pay-per-token on FriendliAI. Try it today ➜
Blackfrost-AI
Available on FriendliAI
Run this model inference on single tenant GPU with unmatched speed and reliability at scale.
Run this model inference with full control and performance in your environment.
Model Details
Model Provider
Blackfrost-AI
Model Tree
moonshotai/Kimi-K2.7-Code
Input Modalities
Output Modalities
Supported Functionality
Kimi K2.7 Code is a coding-focused agentic model built upon Kimi K2.6. With substantial improvements on real-world long-horizon coding tasks, it strengthens end-to-end task completion across complex software engineering workflows while improving token efficiency, reducing thinking-token usage by approximately 30% compared with Kimi K2.6.
| Architecture | Mixture-of-Experts (MoE) |
| Total Parameters | 1T |
| Activated Parameters | 32B |
| Number of Layers (Dense layer included) | 61 |
| Number of Dense Layers | 1 |
| Attention Hidden Dimension | 7168 |
| MoE Hidden Dimension (per Expert) | 2048 |
| Number of Attention Heads | 64 |
| Number of Experts | 384 |
| Selected Experts per Token | 8 |
| Number of Shared Experts | 1 |
| Vocabulary Size | 160K |
| Context Length | 256K |
| Attention Mechanism | MLA |
| Activation Function | SwiGLU |
| Vision Encoder | MoonViT |
| Parameters of Vision Encoder | 400M |
| Weight Precision (this repo) | BF16 (dequantized from native INT4) |
Evaluations are those of the original
moonshotai/Kimi-K2.7-Code. Dequantization to BF16 does not change model behavior; if anything it removes INT4 rounding.
The original Kimi-K2.7-Code is released in native INT4 (pack-quantized), using the same method as Kimi-K2-Thinking.
This repository is the full BF16 dequantization of that release. The INT4 packed routed-expert weights were unpacked and dequantized to bfloat16; attention, shared experts, dense layer 0, embeddings, lm_head, and the vision tower (already BF16 in the original) are carried through unchanged. config.json has quantization_config removed and dtype: bfloat16, so it loads as a standard BF16 model — no compressed-tensors needed.
Use these weights as a full-precision starting point to re-quantize (GGUF, NVFP4, FP8, AWQ, …) or to fine-tune.
Kimi-K2.7-Code has the same architecture as Kimi-K2.5/Kimi-K2.6, and the deployment method can be directly reused. Recommended inference engines:
This BF16 checkpoint loads with trust_remote_code=True at dtype=torch.bfloat16. Note it is ~2 TB — plan tensor-parallel/offload accordingly. For the original INT4 release and its official deployment guide, see moonshotai/Kimi-K2.7-Code.
Usage is identical to the original model (forced thinking + preserve_thinking). See the original model card for chat-completion, vision, preserve-thinking, and tool-call examples. The chat template and tokenizer in this repo are the originals, unchanged.
Released under the Modified MIT License, the same license as the original model weights. Copyright © 2026 Moonshot AI. This BF16 conversion by Blackfrost-AI is distributed under the same terms.
See THIRD PARTY NOTICES.
Full credit to Moonshot AI for creating and open-sourcing Kimi-K2.7-Code. This repository only redistributes a dequantized copy of their weights. For questions about the original model, contact support@moonshot.ai.