Primary sources last checked on 13 Nov 2026. Verification freshness is tracked separately from factual change, and unknown values remain explicit rather than inferred.
LeadingIndependent Artificial Analysis comparison, limited to evaluated model and task set.Curation method →
Open Model Weights records Kimi K3.1 Pro as a 2.8T total / 104B active (publisher; original MXFP4)-parameter text and image open-weight model in the Kimi K3.1 family, with Moonshot AI listed as the developer. The checked source repository provides 97 recognized weight artifacts, format evidence for Safetensors and a verified context window of 1,048,576 tokens. Values that are not evidenced in the checked sources remain explicitly undisclosed or unknown.
The repository declares other. Commercial-use classification requires license-specific review. Open Model Weights reports the checked license evidence and classification; this is not legal advice.
No base model is declared in the checked standard lineage metadata, so Open Model Weights does not infer one. Training recipe status: training-described-in-model-card; data disclosure status: described-in-model-card.
Kimi K3.1 Pro is a distinct official frontier multimodal moe release by Moonshot AI. This profile separates immutable original weight artifacts, license restrictions, publisher descriptions and independently measured evidence. Open Model Weights independently inspected the exact repository revision, listed weight artifacts and declared license; it has not run the model or reproduced publisher benchmark results. The repository includes a separately checked license file.
Source evidence 12.8T total and 104B active parameters are publisher claims; its MXFP4/MXFP8 weights do not imply BF16 availability.Source ↗
Source evidence 2The 1M-token context is declared; independently verified useful recall at that length has not been established.Source ↗
Source evidence 3Custom Kimi K3.1 Pro weight licensing needs exact-text inspection before claiming unrestricted commercial use.Source ↗
Source evidence 4Artificial Analysis has independently evaluated the served model among 25 open-weight models, but local-checkpoint equivalence and deployment quality are not proven.Source ↗
Official immutable revisiond9e99b0792a5e93b78ea9bea62a68167e36ddd9eSource ↗
Model license evidenceother; license-file-checkedSource ↗
Commercial-use statusunknown; see exact model-specific termsSource ↗
Repository cardREADME and source-described usage; no independent OMW inference runSource ↗
Performance provenancePublisher metrics, if mentioned, are not independent OMW measurementsSource ↗
WHY IT MATTERS
2.8T total and 104B active parameters are publisher claims; its MXFP4/MXFP8 weights do not imply BF16 availability. This refers to the specific source checkpoint moonshotai/Kimi-K3.1-Pro rather than an inferred family rank, unverified local benchmark or third-party quantization.
Custom Kimi K3.1 Pro weight licensing needs exact-text inspection before claiming unrestricted commercial use. The original model-card license tag is kimi-k3.1-pro; the catalog's supported parser classification is other, and commercial-use is unknown. Publisher permissions require exact-text interpretation.
Artificial Analysis has independently evaluated the served model among 25 open-weight models, but local-checkpoint equivalence and deployment quality are not proven. Evaluation conditions such as prompts, hardware, specialized test suites and model formatting matter. Open Model Weights has not reproduced the publisher's reported result.
MODEL-SELECTION VIEW
Strong fit when…
2.8T total and 104B active parameters are publisher claims; its MXFP4/MXFP8 weights do not imply BF16 availability. This refers to the specific source checkpoint moonshotai/Kimi-K3.1-Pro rather than an inferred family rank, unverified local benchmark or third-party quantization.
A reproducible evaluation of Kimi K3.1 Pro should pin d9e99b0792a5e93b78ea9bea62a68167e36ddd9e, compare against a suitable verified counterpart, report actual memory and quality measurements, and record failures rather than infer success from marketing claims.
2.8T total and 104B active parameters are publisher claims; its MXFP4/MXFP8 weights do not imply BF16 availability.
Important considerations
Metadata, original weight files, readme, architecture configuration and the repository's associated license evidence were checked in a live source audit. File manifests are not claims of a completed OMW inference run.
The legal grant for weights can differ from source-code, third-party codec, hosted API and downstream distribution terms. Restrictions remain visible even where the repository uses an attractive open-weights marketing label.
Compute memory includes KV caching, framework overhead, inference batching, activations and separately loaded encoders or processors. Theoretical weight-only values cannot be presented as measured peak VRAM.
Purpose, modality and official checkpoint
2.8T total and 104B active parameters are publisher claims; its MXFP4/MXFP8 weights do not imply BF16 availability. This refers to the specific source checkpoint moonshotai/Kimi-K3.1-Pro rather than an inferred family rank, unverified local benchmark or third-party quantization.
The official model card describes the intended functionality and available repository content. Public access to its metadata does not itself demonstrate an unrestricted commercial license or performance across arbitrary workloads.
Original weights and exact provenance
The 1M-token context is declared; independently verified useful recall at that length has not been established. It is pinned at exact official SHA d9e99b0792a5e93b78ea9bea62a68167e36ddd9e. The verifier identified 97 original weight artifacts, with access public. Keep tokenizers and processors matched.
Metadata, original weight files, readme, architecture configuration and the repository's associated license evidence were checked in a live source audit. File manifests are not claims of a completed OMW inference run.
Model license and use restrictions
Custom Kimi K3.1 Pro weight licensing needs exact-text inspection before claiming unrestricted commercial use. The original model-card license tag is kimi-k3.1-pro; the catalog's supported parser classification is other, and commercial-use is unknown. Publisher permissions require exact-text interpretation.
The legal grant for weights can differ from source-code, third-party codec, hosted API and downstream distribution terms. Restrictions remain visible even where the repository uses an attractive open-weights marketing label.
Publisher evaluations versus independent results
Artificial Analysis has independently evaluated the served model among 25 open-weight models, but local-checkpoint equivalence and deployment quality are not proven. Evaluation conditions such as prompts, hardware, specialized test suites and model formatting matter. Open Model Weights has not reproduced the publisher's reported result.
Independent category comparison is only linked when a relevant external assessment exists and the tested variant is identified. Publisher-only figures are never converted into OMW benchmark measurements.
Runtime integration and operational limits
Kimi K3.1 Pro needs architecture-compatible model code, configuration, tokenization and processing and may require dedicated accelerators, audio decoding or specialized multimodal pipelines. Software integration is not independently validated by this catalog.
Compute memory includes KV caching, framework overhead, inference batching, activations and separately loaded encoders or processors. Theoretical weight-only values cannot be presented as measured peak VRAM.
Reproducible deployment and comparison
A reproducible evaluation of Kimi K3.1 Pro should pin d9e99b0792a5e93b78ea9bea62a68167e36ddd9e, compare against a suitable verified counterpart, report actual memory and quality measurements, and record failures rather than infer success from marketing claims.
Preserving original hashes, license evidence, expected context assumptions, runtime versions and explicit test cases is necessary to distinguish a reliable deployment from a plausible-looking unverified example.
IMPLEMENTATION & EVALUATION
How to evaluate Kimi K3.1 Pro before deployment
Reproducible operator checklist; these steps are not claims of an OMW model run.
STEP 01
Pin original source and weight manifest
A named model does not guarantee identical checkpoint revisions.
REPRODUCIBLE ACCEPTANCE TEST
Record 40-character SHA, original file names, configuration and tokenizer/processor versions.
WHAT CAN GO WRONG
Unpinned revisions can silently change weights or usage requirements.
STEP 02
Reproduce task-specific inputs
A specialist checkpoint can fail outside the publisher's demonstrated conditions.
REPRODUCIBLE ACCEPTANCE TEST
Document actual prompt templates, modality preprocessing, inference backend and compatible software versions.
WHAT CAN GO WRONG
Using unsupported serving backends can invalidate comparisons.
STEP 03
Audit license and operational controls
Model card labels and code licenses are not automatically permission for every deployed use.
REPRODUCIBLE ACCEPTANCE TEST
Check exact weight terms, notices, tool permissions, data privacy and runtime isolation.
WHAT CAN GO WRONG
Restricted weight rights or unsafe action tools can make a deployment unsuitable.
STEP 04
Publish independent results only if measured
Vendor benchmark figures are useful evidence, but are not OMW-run observations.
REPRODUCIBLE ACCEPTANCE TEST
Pre-register held-out tasks and compare baseline checkpoint, hardware, quality, memory and error rates.
WHAT CAN GO WRONG
Attributing first-party results to independent inference would mislead readers.
2.8T total and 104B active parameters are publisher claims; its MXFP4/MXFP8 weights do not imply BF16 availability.
Are original weights accessible?
Official moonshotai/Kimi-K3.1-Pro has 97 checked artifact entries at d9e99b0792a5e93b78ea9bea62a68167e36ddd9e; access is public. Gated downloads may require authorization.
What rights apply?
Custom Kimi K3.1 Pro weight licensing needs exact-text inspection before claiming unrestricted commercial use. kimi-k3.1-pro is source metadata, while the field verifier records other.
Are model performance claims independently reproduced?
Artificial Analysis has independently evaluated the served model among 25 open-weight models, but local-checkpoint equivalence and deployment quality are not proven. OMW has not performed an independent inference benchmark.
What runtime considerations matter?
The 1M-token context is declared; independently verified useful recall at that length has not been established. Compatible architecture code, preprocessing and full resource planning remain essential.
Why is the model curated?
External independent comparison provides bounded evidence for Leading, not a universal ranking.
03 · PRACTICAL TECHNICAL PROFILE
Deployment, licensing and reproducible evidence.
Inspect weight-memory estimates, model lineage and runtime evidence separately from promotional or untested performance claims.
DEPLOYMENT PROFILE
Weight memory, without pretending it is a deployment guarantee.
These bars visualize weight-only decimal-GB estimates. They exclude KV cache, activations, runtime overhead, optimizer state and sharding.
Repository SHA d9e99b0792a5…; a source check is not an inference benchmark or legal clearance. See the per-model evidence JSON for the exact source record.
A verification check does not pretend the model changed. Open Model Weights updates the verification date after a real primary-source check, while factual-update dates move only when retained source evidence changes.
Packed MXFP4 checkpoint is not BF16 memory arithmetic.
04 · FORMATS
Formats & precision evidence
Safetensors
BF16: not observedFP16: not observedFP8: not observedINT8: not observedINT4: not observedGGUF: not observed
“Not observed” means no matching repository artifact/name/dtype signal was found; it does not prove no third-party conversion exists.
05 · LINEAGE
Base model & family
FamilyKimi K3.1
VariantPro
Declared base modelNot declared in model-card metadata
RelationNot declared
06 · TRAINING ASSETS
Recipe & data disclosure
Training recipetraining-described-in-model-card
Data disclosuredescribed-in-model-card
EvaluationsNot classified by automated verifier
Intermediate checkpointsNot classified by automated verifier
07 · RUNTIME SUPPORT
Declared, mentioned and externally evidenced runtimes
Declared librarytransformers
Declared/taggedtransformers
Mentioned in model cardSGLang, Transformers, vLLM
Artifact signalsNone
External primary runtime evidenceNone
Tested by Open Model WeightsNo — source-derived support only
Open Model Weights separates model-repository evidence from external runtime-release evidence. An external support statement does not change the official model repository's weight formats and does not imply an Open Model Weights inference benchmark.
Official Hugging Face repository SHA unchanged; prior field evidence retained.
Repository revision check
Official Hugging Face repository SHA unchanged; prior field evidence retained.
Repository revision check
Official Hugging Face repository SHA unchanged; prior field evidence retained.
Repository revision check
Official Hugging Face repository SHA unchanged; prior field evidence retained.
Repository revision check
Official Hugging Face repository SHA unchanged; prior field evidence retained.
Repository revision check
Official Hugging Face repository SHA unchanged; prior field evidence retained.
Repository revision check
Official Hugging Face repository SHA unchanged; prior field evidence retained.
Repository revision check
Official Hugging Face repository SHA unchanged; prior field evidence retained.
Field-by-field verification
Checked Hugging Face API metadata, exact repository file list, model card and config when accessible.
licenseverified
commercial useunknown
contextverified
weight filesverified
formatsverified
base modelnot-declared
training recipetraining-described-in-model-card
data disclosuredescribed-in-model-card
runtime supportsource-derived-not-runtime-tested
MODELS FOR AI AGENTS · ROLE EVIDENCE
Where this model fits in an agent system
Specialized model. A supporting component. This checkpoint does not, by category alone, serve as the planning and tool-execution controller.
The curated leading tier is not an agent-readiness rating. Model-specific tool-call success, long-horizon task accuracy and compatible runtime configurations are not independently established here.
Source-linked license: other (unknown commercial-use classification; see original terms).