Home / Models / K2-Horizon-375B-A23B

INSTITUTE OF FOUNDATION MODELS (IFM) · FIELD VERIFIED

K2-Horizon-375B-A23B

Checked against the listed source repository on 2026-10-07. Unknown values are left explicit rather than inferred.

LeadingCurrent independently benchmarked frontier-relevant open-weight model with a first-party source-verified canonical record.Curation method →

Source repository ↗Compare this model →History & diff →Machine-readable JSON ↗Field-verified 2026-10-07

Evidence statusSource-first verified

Repository revision checked2026-10-07

Full field verification2026-10-07

Source policy →Methodology →

DIRECT ANSWER · FEATURED MODEL BRIEFING

K2-Horizon-375B-A23B in plain English.

Source-grounded editorial layer · reviewed 2026-10-07

K2-Horizon-375B-A23B is IFM’s flagship K2-Horizon sparse Mixture-of-Experts model. The publisher names the checkpoint for a 375B-total / 23B-active design, while Open Model Weights observes 379.17B parameters in the released Safetensors metadata. The checked config verifies a 524,288-token context window, and the official repository declares Apache 2.0 while linking released pretraining and midtraining datasets.

Verified parameter metadata379.17B parameters observed in the published Safetensors metadataSource ↗

Publisher model designationK2-Horizon-375B-A23B; publisher describes 375B total / 23B activeSource ↗

Verified context524,288 tokensSource ↗

LicenseApache 2.0 declared in official repository metadataSource ↗

Training disclosureOfficial metadata declares pretraining and midtraining datasetsSource ↗

Observed weights62 recognized Safetensors artifacts in the checked repository revisionSource ↗

WHY IT MATTERS

K2-Horizon is unusually useful for evidence-focused model research because the first-party release exposes more than final weights: it links training data, training code and staged model-development information.

Its sparse architecture combines hundreds of billions of stored parameters with a much smaller publisher-stated active parameter count per token and a verified 512K context window.

The Leading badge reflects current independent benchmark relevance only. OMW keeps benchmark curation separate from source-verified facts such as weights, license, context and training disclosures.

MODEL-SELECTION VIEW

Strong fit when…

Long-context text and agent research where a verified 524,288-token configuration is relevant.
Researchers who value first-party training disclosures and released dataset links in addition to model weights.
Teams that prefer a standard Apache 2.0 declaration for the published model artifacts.

Important considerations

The checkpoint remains a large serving target even though the publisher describes only a subset of parameters as active per token.
The name’s 375B figure and the 379.17B Safetensors parameter total describe related but not identical quantities; OMW shows both rather than silently forcing them to match.
Publisher benchmark tables are evidence of publisher claims, not an OMW universal quality score.

Sparse MoE architecture and 512K context

The checked configuration exposes a 524,288-token maximum position setting. The publisher describes K2-Horizon-375B-A23B as a sparse Mixture-of-Experts checkpoint with 375B total and roughly 23B active parameters, while the repository’s Safetensors metadata yields 379.17B published parameters.

Showing both values is intentional. OMW distinguishes publisher architecture naming from parameters observed in the actual released weight metadata instead of hiding the discrepancy.

Training provenance and data disclosure

The current official metadata declares IFM/K2-Horizon-Pretrain-Data and IFM/K2-Horizon-Midtrain-Data. The model card also describes training artifacts and development stages. OMW therefore records both the training recipe and dataset disclosure as present in the checked sources.

This level of disclosure makes K2-Horizon especially relevant for users comparing open-weight models on reproducibility and provenance, not only benchmark performance.

License, weights and runtime evidence

The official repository declares Apache 2.0. OMW verified 62 recognized Safetensors artifacts and stores the exact file list together with the repository SHA.

Transformers is declared by the repository, while the model card also mentions NeMo, vLLM and SGLang. OMW exposes those as runtime signals without claiming independent execution testing.

PRIMARY & INDEPENDENT SOURCES

Official model repositoryInspect ↗Official model cardInspect ↗Official config.jsonInspect ↗Official model APIInspect ↗Independent open-weight comparisonInspect ↗

FAQ

Questions people ask about K2-Horizon-375B-A23B.

What is K2-Horizon-375B-A23B?

It is IFM’s flagship K2-Horizon sparse Mixture-of-Experts open-weight checkpoint. The publisher describes a 375B-total / 23B-active architecture.

Why does Open Model Weights show about 379.17B parameters for K2-Horizon?

The 375B figure is the publisher’s model designation and architecture description. OMW independently sums the parameters reported by the released Safetensors metadata and observes about 379.17B.

What is the K2-Horizon context window?

The checked configuration declares 524,288 tokens, commonly described as 512K context.

What license does K2-Horizon-375B-A23B use?

The official repository metadata declares Apache 2.0, and OMW classifies commercial use as allowed under Apache 2.0 conditions.

Does K2-Horizon disclose training data?

Yes. The current official metadata declares separate K2-Horizon pretraining and midtraining datasets, and the model card describes training artifacts.

Are the K2-Horizon weights downloadable?

Yes. OMW verified a public first-party repository containing 62 recognized Safetensors artifacts in the checked revision.

MODEL INTELLIGENCE · SOURCE-FIRST

What is K2-Horizon-375B-A23B?

OMW evidence review7 Oct 2026

Open Model Weights records K2-Horizon-375B-A23B as a 379.167B-parameter text open-weight model in the K2-Horizon family, with Institute of Foundation Models (IFM) listed as the developer. The checked source repository provides 62 recognized weight artifacts, format evidence for Safetensors and a verified context window of 524,288 tokens. Values that are not evidenced in the checked sources remain explicitly undisclosed or unknown.

The repository declares apache-2.0. Commercial use allowed under apache-2.0 conditions. 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: dataset-metadata-declared.

Source repository ↗Model card ↗Configuration ↗

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.

WEIGHT-ONLY MEMORY

BF16 / FP16~758.3 GB

FP8 / INT8~379.2 GB

INT4~189.6 GB

FIELD EVIDENCE MAP

WeightsVerified

LicenseVerified

Commercial useAllowed

ContextVerified

FormatsVerified

LineageNot declared

Training recipeTraining described in model card

Training dataDataset metadata declared

RuntimeSource-derived · not runtime-tested

LICENSE & USE

apache-2.0

Commercial use allowed under apache-2.0 conditions

Declared licenseapache-2.0

ClassificationAllowed

VerificationDeclared metadata checked

This section summarizes the checked repository evidence. It is not legal advice.

LINEAGE & PROVENANCE

What the checked sources disclose — and what they do not.

Open Model Weights does not infer a hidden base model or training corpus when the checked sources do not declare one.

LINEAGE

BASE / PARENTNot declared in model-card metadataUnknown

not inferred

CURRENT RECORDK2-Horizon-375B-A23BK2-Horizon

TRAINING DISCLOSURE

Training recipeTraining described in model card

Training dataDataset metadata declared

FamilyK2-Horizon

RUNTIME SIGNALS

Compatibility claims stay attached to their evidence level.

Declared and model-card-mentioned runtimes are discovery signals. They are not treated as Open Model Weights runtime tests unless explicitly measured.

Declaredtransformers

MentionedNeMo

MentionedSGLang

MentionedTransformers

MentionedvLLM

OMW runtime testNot performed

EVIDENCE TRAIL

Sources and observed history.

This is where the article remains auditable: source links on one side, the registry's observed change history on the other.

PrimaryModel Repositoryhttps://huggingface.co/IFM/K2-Horizon-375B-A23BInspect source ↗PrimaryModel Apihttps://huggingface.co/api/models/IFM/K2-Horizon-375B-A23BInspect source ↗PrimaryModel Cardhttps://huggingface.co/IFM/K2-Horizon-375B-A23B/blob/main/README.mdInspect source ↗PrimaryConfighttps://huggingface.co/IFM/K2-Horizon-375B-A23B/blob/main/config.jsonInspect source ↗

2026-10-07

Repository revision check

Official Hugging Face repository SHA unchanged; prior field evidence retained.

2026-10-07

Field-by-field verification

Checked Hugging Face API metadata, exact repository file list, model card and config when accessible.

01 · WEIGHTS

Exact repository artifacts

Weightsverified-available

Accesspublic

RepositoryIFM/K2-Horizon-375B-A23B

Recognized weight files62

Show all 62 exact weight files

HF API file listing ↗

02 · LICENSE

License & commercial use

Declared licenseapache-2.0

License IDapache-2.0

Commercial useCommercial use allowed under apache-2.0 conditions

Repository license filesNone observed

Classification describes the checked license terms; it is not legal advice.

03 · HARDWARE

Weight-memory estimate

BF16 / FP16~758.3 GB

FP8 / INT8~379.2 GB

INT4~189.6 GB

Weight-only decimal-GB estimate from total parameter count; excludes KV cache, activations, runtime overhead, optimizer state and sharding.

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

FamilyK2-Horizon

Variant375B-A23B

Declared base modelNot declared in model-card metadata

RelationNot declared

06 · TRAINING ASSETS

Recipe & data disclosure

Training recipetraining-described-in-model-card

Data disclosuredataset-metadata-declared

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 cardNeMo, SGLang, 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.

08 · SOURCES

Field evidence

Primary / declared sources4

Repository revisionbd1606f5332f

Source policyHow evidence is classified →

model_repositoryhttps://huggingface.co/IFM/K2-Horizon-375B-A23B

model_apihttps://huggingface.co/api/models/IFM/K2-Horizon-375B-A23B

model_cardhttps://huggingface.co/IFM/K2-Horizon-375B-A23B/blob/main/README.md

confighttps://huggingface.co/IFM/K2-Horizon-375B-A23B/blob/main/config.json

09 · VERIFICATION HISTORY

Checked, not assumed

2026-10-07

Repository revision check

Official Hugging Face repository SHA unchanged; prior field evidence retained.

2026-10-07

Field-by-field verification

Checked Hugging Face API metadata, exact repository file list, model card and config when accessible.

licenseverified

commercial useclassified

contextverified

weight filesverified

formatsverified

base modelnot-declared

training recipetraining-described-in-model-card

data disclosuredataset-metadata-declared

runtime supportsource-derived-not-runtime-tested