K2 Horizon 3.7B

Fiche modèle · Institute of Foundation Models · k2-horizon-3.7b

K2 Horizon 3.7B is the small dense member of IFM's (MBZUAI) fully open K2 Horizon family: a 3.7B-core decoder-only reasoning model (36 layers; about 5.1B parameters in the released safetensors including its 250K-token vocabulary embeddings) pretrained on 22.9T tokens, with a native 524,288-token (512K) context window from midtraining onward. Text in, text out, with per-request reasoning effort (high recommended) and JSON/XML tool-call formats. Released 2026-09-03 under Apache 2.0; on 2026-10-01 the main branch was updated to an anti-doom-loop (FTPO) checkpoint, and the model-card scores were refreshed on 2026-10-28 (they differ from the launch-blog figures). BF16 weights plus a GGUF build, intermediate checkpoints, training data, code, and logs. No priced provider is cataloged yet.

Identité

OrganisationInstitute of Foundation Models
Pays—
Famille—
Typellm
Date de sortie2026-09-03
Licenceapache 2.0
Licence commercialeoui
Licence open sourceoui
Poids ouvertsoui
Auto-hébergement (DGX Spark)oui — ≤ 107 Md (Q8/FP8)
Modalités—
Fenêtre de contexte524 288
KV cache / token (Q8)73.7 ko/token
Paramètres totaux3.7 B
Paramètres actifs (session)5.06 B
Source des paramètres actifsactifs = total (modèle dense, config HF (IFM/K2-Horizon-3.7B)) ; total : config Hugging Face (IFM/K2-Horizon-3.7B)
Curation des paramètres actifs2026-11-04
Prix par session—
Tokens / tâche (Data Players)1.2 M
Tokens / tâche (Terminal-Bench)83.3 M (est.)
Ratio Terminal-Bench / Data Players69.4×
Coût / tâche (Terminal-Bench)—
Tokens d'entraînement22 900 000 000 000
Coupure de connaissance—
Créé le2026-10-29T18:18:04.875068Z
Mis à jour le2026-11-04T06:04:48.146265Z
SourceLLM Stats
URLhttps://llm-stats.com/models/k2-horizon-3.7b

Prix par fournisseur

Statut du modèle (offres + auto-hébergement) : partiel — auto-hébergement DGX Spark (≤ 107 Md).

Toutes sources confondues : catalogue llm-stats, routeurs EUrouter/Cortecs et catalogues directs (OVHcloud, Scaleway, Infomaniak). Les prix sont convertis en USD pour la comparaison ; le prix publié et le taux BCE utilisé sont dans l'infobulle. Les prix inconnus restent indiqués « — ».

Aucune donnée.

Souveraineté qualifiée par offre (modèle × fournisseur) ; survolez le badge pour les critères.

Souveraineté

Scores par catégorie (top_scores)

CatégorieScore
math0.40
physics0.50
reasoning0.40
frontend_development0.70
general0.50
agents0.40
biology0.50
chemistry0.50
code0.40
tool_calling0.40
vision0.10

Scores par benchmark

BenchmarkCatégorieScoreMaxNormalisé (API)VérifiéAuto-déclaréDaté du
bfcl-v4tool_calling0.651.000.65nonoui2026-11-04T06:04:48.223087Z
gpqareasoning0.691.000.69nonoui2026-11-04T06:04:48.173072Z
hmmt-feb-26math0.751.000.75nonoui2026-11-04T06:04:48.153107Z
humanity's-last-examreasoning0.141.000.14nonoui2026-11-04T06:04:48.183135Z
scicodereasoning0.261.000.26nonoui2026-11-04T06:04:48.193063Z
swe-bench-verifiedreasoning0.661.000.66nonoui2026-11-04T06:04:48.163017Z
tau3-bankingagents0.181.000.18nonoui2026-11-04T06:04:48.212932Z
terminal-bench-2.1reasoning0.401.000.40nonoui2026-11-04T06:04:48.203116Z

Classements

CatégorieRangScoreNote conservatriceBenchmarks évaluésMéthodeDaté duProvenance
agents152—9.723trueskill2026-11-02model_history
code220—6.393trueskill2026-11-02model_history

Lecture : Note conservatrice = borne basse TrueSkill (niveau estimé − marge d’incertitude) ; Benchmarks évalués = nombre de benchmarks de la catégorie pris en compte. À performances égales, un modèle évalué sur moins de benchmarks obtient une note plus basse : la note ne se déduit pas d’un seul score. Provenance : rankings = top 50 de /rankings ; model_history = historique du modèle (/models/{id}/history), utilisé pour les rangs au-delà du top 50.

Données

Dates de curation (récupération depuis l'API) : models 2026-11-04 · benchmarks 2026-11-04 · scores 2026-11-04 · updates 2026-11-04 · account 2026-11-04 · model_history 2026-11-04 · rankings 2026-11-04