K2 Horizon 375B-A23B

Institute of Foundation Models

K2 Horizon 375B-A23B is the flagship of the Institute of Foundation Models' (MBZUAI) fully open K2 Horizon family: a sparse Mixture-of-Experts reasoning model with 375B total and 23B active parameters per token (61 layers, 192 routed experts with 8 active plus 1 shared expert, 250K vocabulary). It was pretrained on about 15.1T tokens in two phases, context-extended to a native 524,288-token (512K) window during midtraining, and post-trained by merging RL experts for knowledge work, instruction following, search, tool use, and reasoning. Text in, text out, with per-request reasoning effort (high recommended) and JSON/XML tool-call formats. It targets agentic tool use, terminal, coding, and long-horizon workflow tasks. Released 2026-09-03 under Apache 2.0 with BF16 weights plus FP8 and NVFP4 builds, intermediate checkpoints, training data, code, and logs. IFM lists Compass, Cerebras, and Nebius as API inference partners; no priced provider is cataloged yet.

Key Specifications

Parameters
375.0B
Context
-
Release Date
September 3, 2026
Average Score
49.7%

Timeline

Key dates in the model's history
Announcement
September 3, 2026
Last Update
October 31, 2026
Today
November 20, 2026

Technical Specifications

Parameters
375.0B
Training Tokens
15.1T tokens
Knowledge Cutoff
-
Family
-
Capabilities
MultimodalZeroEval

Benchmark Results

Model performance metrics across various tests and benchmarks

Reasoning

Logical reasoning and analysis
GPQA
IFM K2-Horizon-375B-A23B Hugging Face model card, final checkpoint; reasoning_effort=high, temperature=1.0, top_p=0.95; GPQA Diamond • Self-reported
87.3%

Other Tests

Specialized benchmarks
AA-LCR
IFM K2-Horizon-375B-A23B Hugging Face model card, final checkpoint; reasoning_effort=high, temperature=1.0, top_p=0.95; long-context reasoning • Self-reported
76.0%
APEX-Agents
IFM K2-Horizon-375B-A23B Hugging Face model card, final checkpoint; reasoning_effort=high, temperature=1.0, top_p=0.95; pass@1 on the text-only-modality subset of tasks • Self-reported
24.8%
AutomationBench
IFM K2-Horizon-375B-A23B Hugging Face model card, final checkpoint; reasoning_effort=high, temperature=1.0, top_p=0.95; Automation Bench public set • Self-reported
25.3%
BrowseComp
IFM K2-Horizon-375B-A23B Hugging Face model card, final checkpoint; reasoning_effort=high, temperature=1.0, top_p=0.95; Discard-all@95k context-management protocol from the DeepSeek-V3.2 report • Self-reported
72.8%
CritPT
IFM K2-Horizon-375B-A23B Hugging Face model card, final checkpoint; reasoning_effort=high, temperature=1.0, top_p=0.95 • Self-reported
8.6%
GDPval-AA
IFM K2-Horizon-375B-A23B Hugging Face model card, final checkpoint; reasoning_effort=high, temperature=1.0, top_p=0.95; Elo rating on real-world professional tasks • Self-reported
48.0%
Humanity's Last Exam
IFM K2-Horizon-375B-A23B Hugging Face model card, final checkpoint; reasoning_effort=high, temperature=1.0, top_p=0.95; without tools • Self-reported
32.0%
MCP-Mark
IFM K2-Horizon-375B-A23B Hugging Face model card, final checkpoint; reasoning_effort=high, temperature=1.0, top_p=0.95 • Self-reported
67.7%
OmniScience
IFM K2-Horizon-375B-A23B Hugging Face model card, final checkpoint; reasoning_effort=high, temperature=1.0, top_p=0.95; OmniScience accuracy • Self-reported
23.0%
OmniScience (non-hallucination rate)
IFM K2-Horizon-375B-A23B Hugging Face model card, final checkpoint; reasoning_effort=high, temperature=1.0, top_p=0.95 • Self-reported
74.7%
SciCode
IFM K2-Horizon-375B-A23B Hugging Face model card, final checkpoint; reasoning_effort=high, temperature=1.0, top_p=0.95 • Self-reported
42.7%
SWE Atlas - Codebase QnA
IFM K2-Horizon-375B-A23B Hugging Face model card, final checkpoint; reasoning_effort=high, temperature=1.0, top_p=0.95; SWE-Atlas-QnA strict (no internet) • Self-reported
48.4%
SWE-Bench Pro
IFM K2-Horizon-375B-A23B Hugging Face model card, final checkpoint; reasoning_effort=high, temperature=1.0, top_p=0.95; strict (no internet) • Self-reported
42.6%
Tau3 Banking
IFM K2-Horizon-375B-A23B Hugging Face model card, final checkpoint; reasoning_effort=high, temperature=1.0, top_p=0.95 • Self-reported
34.0%
Terminal-Bench 2.1
IFM K2-Horizon-375B-A23B Hugging Face model card, final checkpoint; reasoning_effort=high, temperature=1.0, top_p=0.95; model card figure; IFM's launch blog notes a reward-hacking audit that revised this to 66.9 • Self-reported
70.2%
Toolathlon-Verified
IFM K2-Horizon-375B-A23B Hugging Face model card, final checkpoint; reasoning_effort=high, temperature=1.0, top_p=0.95; Toolathlon Verified • Self-reported
65.3%
WildClawBench
IFM K2-Horizon-375B-A23B Hugging Face model card, final checkpoint; reasoning_effort=high, temperature=1.0, top_p=0.95; subset of English text-only-modality tasks • Self-reported
50.9%

License & Metadata

License
apache_2_0
Announcement Date
September 3, 2026
Last Updated
October 31, 2026

Similar Models

All Models

Recommendations are based on similarity of characteristics: developer organization, multimodality, parameter size, and benchmark performance. Choose a model to compare or go to the full catalog to browse all available AI models.