Atlas

Models

← All models

K2-V2-Instruct

LLM360 · 2025-12-05 · 72.6B parameters

K2-V2-Instruct is LLM360's instruction-tuned 70B K2-V2 model exposing low, medium, and high reasoning-effort settings through the chat template. It inherits K2-V2's 14.969T pretraining and mid-training tokens and adds three supervised fine-tuning epochs of approximately 17.4B tokens each.

Benchmark scores

BenchmarkScore
AA-LCR33.3
AA-LCR v1.135.0
AA-Omniscience Index-47.3
AIME 2025 (Artificial Analysis avg@10)78.3
Artificial Analysis Omniscience Accuracy18.5
Artificial Analysis Omniscience Hallucination Rate76.0
Artificial Analysis Openness Index88.9
Artificial Analysis Progress Index v1214
Capability131.8
CritPt (avg@5)0.0
Global MMLU Lite78.6
GPQA Diamond (Artificial Analysis avg@5)68.1
Humanity's Last Exam (Text-Only)10.5
IFBench (Artificial Analysis avg@5)60.1
LiveCodeBench (Artificial Analysis)69.4
MMLU Pro78.6
RegretBench (goal success)39.1
RegretBench (normalized regret)1.0
SciCode (Artificial Analysis avg@3)28.6
Terminal-Bench Hard9.8
τ²-bench (Telecom)27.8
Loading Atlas data…