Kimi-K3.1-Pro-Coder

Kimi-K3.1-Pro-Coder

Runs a heavily compressed copy of the 2.8T-parameter Kimi-K3.1-Pro, tuned for coding and math. Made by Grawnch with a three-tier mixed-bit scheme, for people with high-end GPU workstations.

Made by Grawnch on Hugging Face. These notes are Mindket's, written from its model page and model card; the model's own files and license are on Hugging Face.

About

What it is: Kimi-K3.1-Pro-Coder is a compressed copy of moonshotai/Kimi-K3.1-Pro, a 2.8T-parameter mixture-of-experts model, made by Grawnch. A three-tier mixed-bit scheme called Strata-Coder was tuned on how often experts fire during coding and math work. The result is a set of 96 shards totaling 774.97 GB, where busy experts keep more bits and rarely used ones are squeezed hardest.

Who it suits: People with high-end hardware who want an enormous coding-focused model in fewer resources, and researchers who study compression of mixture-of-experts models.

Industry | Who uses it | How they use it | Example
Software | Developers with GPU workstations | Run a huge coding model with weights split over GPU memory, system RAM and NVMe | Solving a LeetCode-style problem
Research | Quantization researchers | Study mixed-bit compression of expert layers | Comparing the card's agreement and perplexity figures
AI platforms | Infrastructure engineers | Test tiered offloading with the Strata engine | Checking throughput on one card

Anyone without a workstation-class GPU can skip it, since the card's tests ran on a card with 102.6 GB of VRAM.

Where to use it:

• The author's Strata engine, a loader that spreads weights over GPU memory, system RAM and an NVMe drive
• The author's dr-triton kernels for fused Triton dequantization
• NVIDIA RTX PRO 6000 Blackwell Server Edition, the card used for the card's tests
• Windows, macOS or Linux with Python, for the download

Use case | Who | What they get back
Coding problems | Developers | Solutions to programming tasks from a coder-tuned model
Math problems | Students and researchers | Integer answers to competition-style questions
Offloading tests | Infrastructure engineers | A look at tiered loading across VRAM, RAM and NVMe

How to set it up:

• Python and the huggingface_hub package
• The Strata engine and the dr-triton kernels, both named in the card
• A large-VRAM GPU, plenty of system RAM and fast NVMe storage

1. Install Python and the huggingface_hub package with the first command below.
2. Download the model files with the second command, which is a huge transfer.
3. Get the Strata engine and the dr-triton kernels from the author.
4. Load the model through Strata so the three storage tiers are used.

pip install -U huggingface_hub
hf download Grawnch/Kimi-K3.1-Pro-Coder

How to use it: The card shows no run command. It points to the Strata engine for loading. That engine keeps the hottest experts in GPU memory and a warm set in pinned system RAM. The cold tail stays on an NVMe drive and is read through mmap.

Tool | What it does
Strata | Loads the model across VRAM, RAM and NVMe with router lookahead
dr-triton | Provides fused Triton kernels that dequantize the weights

Example: You give it a competitive programming problem of the kind in the card's LiveCodeBench set. On the card's rolling-window set of 30 problems, the base model and this copy each solved 29 of them.

For newcomers and experts: This is an advanced model that needs special loading software and a lot of hardware, so a first step is to read the card and note which pieces you would need. Skip the download until you have them.

The base model has 93 layers with 896 routed experts per layer. Non-routed parts, including attention and the vision tower, stay in BF16 or F32. The routed experts are split by use into three tiers of 15, 30 and 55 percent. The card reports about 0.02 mean KL divergence from the base and a perplexity rise of roughly one to three percent.

Rank experts by coding and math use
Assign each expert to a tier
Compress cold experts with fewer bits
Load hot, warm and cold tiers apart
Generate with Strata

Good to know:

• License: Apache 2.0
• Cost to run: no paid service named, but you need a high-end GPU, system RAM and fast NVMe storage
• Data it reaches: the download fetches files from Hugging Face, and the Strata and dr-triton projects are separate downloads
• The card's tests are the author's own and use only 30 problems per set, so treat them as hints
• The card gives 774.97 GB for the full set, while Hugging Face's record lists about 1,270 GB

Title: Size by tier in GB
Non-routed: 114.76 GB
Tier 1 hot: 160.58 GB
Tier 2 warm: 254.53 GB
Tier 3 cold: 244.95 GB

Title: Effective bits per weight by expert tier
Tier 1 hot: 2.81 bpw
Tier 2 warm: 2.56 bpw
Tier 3 cold: 1.44 bpw

On Hugging Face, 13 November 2026

• Task: image text to text
• Parameters: 1.3T
• Weights: about 1,270 GB in U8
• Formats: Safetensors
• Downloads in the last 30 days: 695
• Likes: 6
• License: apache-2.0
• Based on: moonshotai/Kimi-K3.1-Pro
• Tags: safetensors, kimi_k31, kimi-k3.1, quantization, mixed-bit, strata, dr-triton, image-text-to-text, 8-bit

Links

• Model: https://huggingface.co/Grawnch/Kimi-K3.1-Pro-Coder
• Files: https://huggingface.co/Grawnch/Kimi-K3.1-Pro-Coder/tree/main

Get the model

Download its files with Hugging Face's command line:

pip install -U huggingface_hub
hf download Grawnch/Kimi-K3.1-Pro-Coder

Then follow the usage in its model card. Nothing in this kit runs by itself.