Kimi K3.1 Pro Coder is a community quantization of Moonshot AI’s Kimi K3.1 Pro, published under the Grawnch account. Its model card lists 96 shards totaling 774.97 GB (721.75 GiB), with expert weights divided among GPU memory, system RAM and NVMe storage. The published performance figures are Grawnch’s own evaluation results.

For background on the base model, see NeoTeo’s earlier coverage of Kimi K3.1 Pro.

A checkpoint’s file size tells you how much data it contains; it doesn’t, by itself, tell you whether its proposed memory layout fits your system.

How the quantization divides the weights

Quantization stores model weights at reduced numerical precision to use less space. Kimi K3.1 Pro Coder uses a mixture-of-experts (MoE) architecture: the model routes each token through selected groups of weights, called experts. Grawnch’s card describes 896 routed experts per MoE layer, with 16 selected per token, and assigns the experts to three tiers based on their stated precision and memory location.

“Bits per weight,” or bpw, is the effective number of bits used to represent each weight. The card says its expert tiers were calibrated using router activation frequencies for coding and mathematical reasoning; it does not describe additional model training.

ComponentPrecisionListed sizeAllocation listed on the card
Non-routed weights16.0 bpw114.76 GBGPU VRAM/shared
Tier 1: hot experts2.81 bpw160.58 GBGPU VRAM
Tier 2: warm experts2.56 bpw254.53 GBPinned system RAM
Tier 3: cold-tail experts1.44 bpw244.95 GBNVMe SSD (mmap)

The three expert tiers account for 15%, 30% and 55% of experts, respectively. The card’s component sizes total 774.82 GB, 0.15 GB less than its stated 774.97 GB total.

The GPU allocation also deserves a close look. The non-routed weights and Tier 1 add up to 275.34 GB. The card names an NVIDIA RTX PRO 6000 Blackwell Server Edition with 102.6 GB of GDDR7 VRAM for its evaluation. The card’s listed allocation includes “shared” memory, but it does not explain how that assignment relates to the GPU’s stated VRAM capacity.

What Grawnch’s evaluation reports

The card describes an evaluation spanning 70 sequences and 22,307 tokens. Its score comparisons use 30 AIME 2025 problems and 30 LiveCodeBench problems from the 2024.10–2025.05 rolling window. Grawnch reports the same number solved by the base and quantized models in each sample:

EvaluationSampleBase modelQuantized model
AIME 202530 problems29/3029/30
LiveCodeBench30 problems29/3029/30

The card also reports top-1 next-token agreement of 95.20% on AIME and 95.06% on LiveCodeBench. Top-5 agreement is 99.91% and 99.92%, respectively. These agreement figures compare the quantized model’s predictions with the base model’s; they are not additional problem-solving scores.

On LiveCodeBench, the card lists perplexity of 2.6911 for the base model and 2.7629 for the quantized one, a reported increase of 2.67%. Perplexity measures how well a model predicts a sequence of tokens; lower values indicate better prediction on that evaluation. Grawnch also labels 36.77 tokens per second as full 93-layer batch throughput. These results describe the author’s stated samples and evaluation setup, rather than performance across other workloads or systems.

What the included license file says

The repository includes a file titled Kimi K3.1-Pro License that names Moonshot AI’s copyright. Under the file’s terms, a qualifying Model-as-a-Service operator must obtain a separate agreement before commercial use if the operator and its affiliates have aggregate revenue above $20 million over any consecutive 12 months.

The file also calls for prominent display of “Kimi K3.1-Pro” in commercial products or services exceeding 100 million monthly active users or $20 million in monthly revenue. It says Sections 2 and 3 do not apply to internal use or use through Moonshot’s official products or certified inference partners. Read the license file included in the repository.